What's new in v0.13.4

Rapid-MLX 0.13.4

Rapid-MLX 0.13.4 makes qualified local models faster under concurrent work, adds an experimental end-to-end video workspace, and turns Community Benchmark into a private-by-default local workspace with explicit sharing. It also improves memory fitting, model discovery, Markdown rendering, and headless operation.

Highlights

Faster qualified Qwen models by default

The four exact qualified Qwen3.5/Qwen3.6/Qwen3.8 artifacts now select their validated MTP preset and continuous scheduler automatically when no speculative-decoding option is supplied. Across mixed four-request cohorts, aggregate throughput improved by 14.1%–30.8%, with no ordinary-pass case becoming a continuous-mode failure. CLI and Server users can restore ordinary decoding with --no-spec-decode; Desktop exposes the same persistent off switch.

Qualified artifact

Aggregate throughput change

Qwen3.5 4B 4-bit

+30.8%

Qwen3.5 9B 4-bit

+23.9%

Qwen3.6 27B 4-bit

+14.1%

Qwen3.8 27B 4-bit

+25.9%

For Qwen3.8 27B, the qualified single-request MTP path also scales with context: measured decode throughput was 1.43× the 0.13.3 ordinary path at 128 tokens of context and 2.34× at 32K. The new continuous scheduler keeps request/cache ownership transactional across admission, cancellation, and dynamic joins.

Local video generation in Desktop

Enable Video Generation under Experimental Features to use a dedicated local workspace with queued jobs, progress, cancellation, restart-safe completed results, and early memory/workload validation. The signed app includes the audited runtime and encoder needed for LTX 2.5 q8 (text/image to video), Wan 2.1 1.3B bf16 (text to video), and CogVideoX-Fun 5B q4 (text to video), without post-install package provisioning. Wan 2.1 is limited to Macs with at least 40 GB unified memory.

Community Benchmark, private by default

rapid-mlx benchmark and the new Desktop Community Benchmark page can plan, run, validate, inspect, and archive text, image, and video measurements on the local Mac. Results remain local unless the user explicitly chooses to share. The consent flow previews the exact destination and payload digest, and a durable receipt prevents accidental duplicate submissions.

Better model and memory decisions

CLI, Server, and Desktop now consume one atomic model registry and recommendation policy. Per-model Metal limits are tuned from the machine's available unified memory, and rejected launches return actionable fit guidance instead of failing deep in model loading. New per-model metrics expose request, token, latency, and throughput totals without mixing aliases.

Desktop quality-of-life improvements

Opt-in persistent memory can retain useful long-term context across conversations and remains editable or removable in Settings.

Mermaid code blocks render as isolated, network-blocked previews.

Dictation overlaps speech-model page-in with the utterance to reduce the perceived first-use delay.

An update card explains whether the installed app is current, behind, or ahead of the public release without forcing an update.

Experimental performance routes

Qwen3.8 Flash-Next gains an opt-in fused single-token GDN kernel. On the qualified M3 Ultra run it improved one complete GDN layer by 26.28% and end-to-end decode from 25.43 to 27.04 tok/s (+6.35%), with bit-identical tokens and recurrent state in the qualification workload. It remains disabled by default while broader hardware evidence is collected. Qwen3.8 27B also adds an explicit experimental DFlash2 route and a conservative prompt-lookup route; unsupported requests continue to fail closed or use the ordinary path.

Reliability and operations

Explicit quantized-KV and paged-cache requests now reject unsupported cache layouts before Ready instead of starting healthy and failing later.

Paged prefix reuse commits hit metrics only after successful reconstruction and rolls back block ownership on failure.

Singleton MTP keeps exclusive scheduler ownership until its request departs, so concurrently arriving plain or tool requests wait instead of aborting both streams with a 503.

A documented LaunchDaemon deployment supports qualified headless macOS servers while keeping the same user-owned cache and service boundaries.

Model-download, GUI attachment, update, video-job, and MTP lifecycle failures have more bounded cleanup and clearer recovery behavior.

Upgrade

pip install -U rapid-mlx

Homebrew follows the PyPI release through its normal autobump process. Desktop users receive the signed and notarized update through the production appcast.

All changes

• chore: bump version to 0.13.4 (711da18)
• fix(mtp): isolate singleton verifier admission (#2964) (f05d8d5)
• feat(benchmark): add consent-gated atomic uploads (#2931) (3b1a553)
• fix: reject explicit quantized KV pre-ready on unsupported families (#2954) (#2959) (9106a36)
• fix(cache): fail closed and transact paged prefix reuse (#2960) (1a5c6bd)
• feat: add local-first Community Benchmark workspace (#2882) (3fe7087)
• feat(video): add LTX 2.5, Wan 2.1, and CogVideoX-Fun to Desktop (#2947) (6096094)
• docs: qualify headless macOS service deployment (#2952) (016a16d)
• catalog: consume atomic recommendation policy (#2871) (87d8825)
• feat(markdown): preview Mermaid diagrams securely (#2812) (86dbcde)
• docs: refresh 0.13.4 large-model benchmarks (#2942) (65a0fbc)
• [codex] Add lightweight desktop update discovery card (#2919) (c9aea88)
• Expose per-model request performance metrics (#2915) (f7ea5bc)
• test(mac): wait for authoritative drop completion before retry (#2938) (ede7158)
• fix(mtp): preserve requests on admission failure (#2934) (bcf8e51)
• feat(catalog): add atomic product model registry (#2857) (97755ce)
• feat(mac): add experimental Video workspace (#2844) (615a8c5)
• fix(doctor): keep runtime probe within five-second contract (#2926) (2a0873f)
• Qualify Qwen3.8-27B DFlash2 as an explicit experimental route (#2920) (abf6a93)
• perf(qwen4): fuse single-token GDN decode (#2918) (4b35ed0)
• Build Desktop video serving foundation (#2810) (63dac4f)
• Fix doctor runtime dependency health (#2865) (7e86e04)
• perf(qwen4): qualify conservative prompt lookup (#2911) (087826f)
• test(mac): retry only unaccepted attachment drags (#2909) (d8b2890)
• fix(mac): overlap dictation model page-in with the utterance itself (#2905) (3024358)
• feat(mac): offer one-click GitHub star via gh CLI (#2772) (01739b3)
• fix(tests): pin prefix-eviction stub to the utilization-ratchet generation (#2907) (d8ffed0)
• feat(proto): define atomic model runtime and benchmark contracts (#2840) (9b50840)
• feat: auto-tune per-model Metal memory limits and return actionable admission errors (#2880) (82b703a)
• fix(ci): use supported recovery for dequeued PRs (#2896) (6c61646)
• fix(ci): follow model memory confirmations in GUI flows (#2826) (612aad4)
• feat(mtp): default verified artifacts to continuous batches (#2875) (1869722)
• fix(ci): allow dequeued PR label recovery (#2893) (813112e)
• fix(ci): recheck stale transcript jump element (#2890) (1faaa7c)
• feat(mtp): adapt continuous batching for Qwen3.8-27B (#2854) (85b5e21)
• fix(ci): requeue dropped heads through ready labels (#2881) (51e6791)
• feat(mtp): add continuous speculative batching foundation (#2842) (644afee)
• feat(mac): persistent memory extraction from completed conversations (#2850) (275b10f)

Install: brew upgrade rapid-mlx or pip install -U rapid-mlx==0.13.4 (or just rapid-mlx upgrade).

Original source

[0.13.4] — 2026-09-02

Rapid-MLX 0.13.4 makes qualified local models faster under concurrent work, adds an experimental end-to-end video workspace, and turns Community Benchmark into a private-by-default local workspace with explicit sharing.

Added

Local video generation in Desktop. An opt-in Video workspace supports LTX 2.5, Wan 2.1, and CogVideoX-Fun with queued jobs, progress, cancellation, restart-safe results, and early memory validation.

Private-by-default Community Benchmark. CLI and Desktop users can run, inspect, and archive reproducible measurements locally; sharing requires an explicit preview and consent, with duplicate submission protection.

Atomic model discovery and recommendations. CLI, Server, and Desktop now consume one validated model registry and recommendation policy.

Opt-in persistent conversation memory, isolated Mermaid previews, per-model request metrics, and clearer in-app update discovery.

Changed

Qualified MTP models are fast by default. Qwen3.5 4B/9B, Qwen3.6 27B, and Qwen3.8 27B automatically use their validated continuous speculative scheduler in CLI, Server, and Desktop. Desktop shows the active setting and keeps a persistent off switch for users who prefer ordinary decoding.

Across mixed four-request cohorts, the qualified paths improved aggregate throughput by 14.1%–30.8%. Qwen3.8 27B decode measured 1.43× the 0.13.3 ordinary path at 128 tokens of context and 2.34× at 32K.

Per-model Metal limits now adapt to available unified memory and failed launches provide actionable fit guidance.

Fixed

Singleton MTP retains exclusive scheduler ownership until its request departs, so concurrently arriving plain or tool requests wait instead of aborting both streams with a 503.

Quantized-KV and paged-cache requests fail closed on unsupported layouts; paged prefix reuse now rolls back ownership cleanly after reconstruction failures.

Model-download, GUI attachment, update, video-job, and MTP lifecycle failures have more bounded cleanup and clearer recovery behavior.

Original source

What's new in v0.13.3

Rapid-MLX 0.13.3

Rapid-MLX 0.13.3 brings GLM-5.3-Flash to the production serving path, makes hybrid multimodal conversations reuse eligible text prefixes, and improves the Desktop workflows around navigation, video generation, streaming output, and failure recovery.

Highlights

GLM-5.3-Flash support

Use glm5.3-flash-4bit from the CLI, OpenAI-compatible server, or Desktop. The release includes the processor and runtime compatibility required by the checkpoint, preserves explicit image-channel layouts, rejects malformed media before model execution, and keeps speculative decoding disabled because its qualification run did not improve throughput.

On a 256 GB M3 Ultra, the qualified 4-bit checkpoint decoded a sustained 512-token response at a median 29.2 tok/s and used 165.4 GB of active MLX memory. The alias therefore requires the 192 GB memory tier. The README's benchmark link records the exact checkpoint revision, request payload, and reproduction command.

Faster repeated work on hybrid vision models

Compatible text-only phases on a hybrid multimodal lane can reuse their text prefix state instead of recomputing it. Cache identity includes the rendered request and model state, while media-bearing requests continue through the validated vision path. This improves repeated background and conversation work without treating an image request as a text-only cache hit.

Desktop workflow improvements

A native command palette makes common actions available from the keyboard. Completed interactions can expose a lightweight feedback entry, while a direct diagnostics shortcut makes failures easier to investigate.

Model pickers remain scoped to the active task instead of leaking selection state across unrelated surfaces.

Streaming Markdown presentation is smoother and remains live when a window is restored away from an attached display.

More resilient video and release workflows

Completed video jobs persist across server restarts, and clients can inspect video capabilities before loading a model. Packaged Desktop builds can recover the engine after a failed post-DMG installation. The merge lane also gains a more reliable native-test bootstrap, dependency-bound evidence, and a faster in-process accessibility journey without reducing required coverage.

Upgrade

pip install -U rapid-mlx

Homebrew follows the PyPI release through its normal autobump process. Desktop users receive the signed and notarized update through the production appcast.

All changes

• fix: fail fast on invalid multimodal runtime (#2862) (6f3f65b)
• test(mac): sink slow-stream-stop, tool-loop-budget, math-rendering, restored-tools into swift test (#2869) (017a36c)
• chore: bump version to 0.13.3 (6e49780)
• docs: refresh recent large-model benchmarks (#2822) (4c03e3d)
• test(mac): sink the message-actions golden journey into swift test (#2855) (0e498ba)
• feat: add production-safe GLM-5.3-Flash support (#2792) (4acffa7)
• fix(ci): preserve mlx guard proof in merge batches (#2833) (75a78cb)
• perf(mllm): reuse text prefixes on hybrid vision lane (#2839) (e986665)
• feat(mac): add post-value feedback entry (#2790) (26b1634)
• feat(mac): add failure diagnostics shortcut (#2786) (a9ddb2c)
• feat(mac): add command palette (#2789) (894f50f)
• fix(ci): settle transcript before fresh-install baseline (#2821) (a0ed91c)
• fix(mac): scope model pickers to the active task (#2785) (68817d9)
• Persist completed video jobs across server restarts (#2794) (ce939da)
• fix(mac): smooth streaming Markdown rendering (#2280) (25a54b6)
• fix(ci): make native XCUITest bootstrap reliable (#2787) (7368ae2)
• Expose pre-serve video model capabilities (#2791) (a4ddc2c)
• fix: add post-DMG engine release recovery (#2332) (1852f33)

Community contributors

• @chen-yijiang — fix(mac): smooth streaming Markdown rendering

Install: brew upgrade rapid-mlx or pip install -U rapid-mlx==0.13.3 (or just rapid-mlx upgrade).

Original source

[0.13.3] — 2026-08-31

Rapid-MLX 0.13.3 adds production-safe GLM-5.3-Flash serving, reuses text prefixes on hybrid vision models, and improves Desktop navigation, video-job recovery, streaming presentation, and failure diagnosis.

Added

GLM-5.3-Flash runs locally through the normal serving path. The new glm5.3-flash-4bit alias includes the processor, quantization, image-layout, and runtime compatibility needed for text and multimodal requests. The qualified 4-bit checkpoint decoded a sustained 512-token response at a median 29.2 tok/s on an M3 Ultra and requires the 192 GB memory tier.

Desktop commands are easier to reach. A native command palette exposes common actions from the keyboard, completed interactions can offer a lightweight feedback entry, and failure diagnostics have a direct shortcut.

Changed

Hybrid vision conversations reuse eligible text prefixes. Repeated text-only phases on the multimodal lane retain compatible prefix state instead of recomputing it, while media-bearing requests continue through their validated vision path.

Large-model performance claims are reproducible. The README now links exact M3 Ultra workloads, checkpoint revisions, memory measurements, and ordinary-versus-MTP results for Qwen3.8 and GLM-5.3-Flash.

GUI regression coverage begins moving deterministic journeys into the Swift test process, shortening the expensive hosted-Mac merge lane without dropping the accessibility contracts those journeys enforce.

Fixed

Completed video jobs survive server restarts, and video capabilities are available before a model is loaded so clients can reject unsupported work early.

Model pickers stay scoped to the active task, streaming Markdown reveals smoothly, and packaged Desktop builds can recover the engine after a failed post-DMG installation.

Native XCUITest bootstrap and merge-queue dependency evidence are more reliable, reducing false release blocks without weakening required checks.

Original source

What's new in v0.13.2

Rapid-MLX 0.13.2 makes long-running local assistants faster and more dependable: Qwen3.8 Flash-Next gains opt-in native MTP and faster long-context prefill, repeated prompts recover their prefix cache, offline speech pulls include their runtime requirements, and Desktop tightens attachment, credential, and web-browsing safety. This stable release also includes the fixes validated after rc1 and replaces the stable Desktop updater feed.

Final release validation

The protected Desktop publication path now promotes the exact signed and notarized candidate bytes, including the canonical DMG and updater payloads, rather than rebuilding after the release tag. (#2775)

On memory-constrained Macs, choosing a photo with a model whose text lane is still usable now explains that text chat remains ready and recommends a lower-memory vision model. The notice clears after the user continues with a text turn, changes model capability, or chooses another attachment path. (#2778)

Highlights

Qwen3.8 Flash-Next gains native MTP

The engine can opt into the checkpoint's one-layer prediction head while the target model verifies every proposal and all recurrent, QSA, and KV state rolls back atomically. The measured fixed-K1 workload accepted 76.41% of proposals. (#2572, #2655)

Context

Serial decode

Native MTP

Change

128

25.17 tok/s

34.85 tok/s

+38.5%

2K

23.64 tok/s

33.53 tok/s

+41.8%

8K

22.82 tok/s

32.20 tok/s

+41.1%

32K

21.16 tok/s

28.82 tok/s

+36.2%

Desktop's normal temperature, top-p, top-k, min-p, and penalty settings can use the lane. Seeded requests and stateful grammar, tool, reasoning, or suppression processors continue on ordinary decoding rather than silently weakening their contract. The measured MTP path used up to 6.6 GB more active memory, and 192 GB remains the recommended hardware tier for this experimental checkpoint.

Faster long-context prefill and reusable prompts

QSA index-cache work is batched across eligible prefills, reducing measured time to first token by 28.9–32.5% while decode speed and cache precision remain unchanged. A completed 32K request can no longer poison the next request in the same process. (#2574, #2596)

Prompt

Previous TTFT

Batched TTFT

Change

2K

3.346 s

2.262 s

-32.4%

8K

13.689 s

9.236 s

-32.5%

32K

62.851 s

44.659 s

-28.9%

Semantic prefix snapshots now follow the exact rendered prompt and preserve the Flash-Next recurrent cache through batching and persistence. A measured 5,288-token warm request reused 5,273 tokens and completed in 0.539 seconds instead of 6.497 seconds; native MTP continued proposing after the hit. (#2588, #2644)

Kokoro is ready before the machine goes offline

rapid-mlx pull now fetches the voice assets and prepares the English G2P requirement as part of the pull transaction. The clean-cache validation downloaded 54 voice files and generated speech with networking disabled. If a runtime requirement is missing, an inference request returns an actionable readiness error instead of trying to download or install software on demand. (#2648, #2664)

A smaller, byte-identical-to-approved Desktop release path

LZMA packaging and dependency-proven pruning reduced the signed and notarized comparison DMG from 187,505,002 to 105,881,983 bytes, a 43.53% reduction. The tradeoff is a slower one-time Finder copy on the measured host: 17.80 seconds instead of 5.94 seconds. Candidate builds now identify their source commit, and protected publication promotes the exact approved DMG and updater payloads instead of rebuilding different bytes after the tag. Tier 1 and Desktop now install the same candidate wheel, with a shared contract matrix keeping text and multimodal request behavior aligned. (#2668, #2450, #2530, #2726)

Desktop safety and workflow corrections

After the first completed exchange, Desktop can derive one short local title without replacing a user rename. Settled text answers can also offer three optional follow-up prompts; malformed, duplicate, wrong-script, or incomplete suggestions stay hidden. (#2698)

A chat message accepts up to four images and 6 MiB of aggregate encoded image data, preserves selection order, and explains whether count or size rejected the remainder. A repeatedly failing image gets one bounded follow-up retry instead of contaminating every later text turn. (#2541, #2585)

Generated-image deletion uses an app-owned confirmation sheet: Keep remains pressable in native GUI automation, Escape is safe, and Return cannot delete an image. (#2387, #2578)

Switching away from a busy model asks first; Cancel preserves the live model and in-flight response. Recommendation, review, and engine admission use one measured footprint for known models, and already-resident models remain usable through the safe replacement path. (#2430, #2543, #2619)

The installer recommends lfm2.5-1b-4bit below 16 GB and qwen3.5-4b-4bit on larger Macs, while preferring an eligible cached model from the same memory tier. (#2426)

Exited server leaders are reaped without blocking a shared worker, and stale dual-stack tcp46 listeners are recognized when Desktop clears a server port. A stopped dictation model reloads when the user next presses the hotkey instead of during foreground activation. (#2562, #2593, #2663)

Tools settings resolve the selected search key's real Keychain state without exposing it, dictation errors no longer guess at an unproven memory cause, and photo guidance names the remedy for the engine's actual serving-lane reason. (#2514, #2523, #2602, #2607)

After 35 successful chats, dictations, or generated images, established Desktop users may see a quiet, nonmodal invitation to visit the project on GitHub. Dismissing it starts a three-day cooldown and a progressively larger local usage threshold; Rapid performs no account or star-status lookup. (#2675)

Engine, API, and CLI reliability

Qwen3.8 27B required and named tool calls use the checkpoint-native format and return OpenAI-compatible JSON arguments. Invalid required arguments fail with a client error instead of appearing executable. (#2660)

Explicit --mllm selection wins over automatic architecture, cache, and runtime fallback. The measured vision-memory floor remains mandatory, and speculative decoding still uses its supported text lane. (#2643, #2669)

Explicit timeout: 0 means the server default, request-local chat_template_kwargs reach the tokenizer, cancellations are distinct from max-length completion, and orphaned streaming reservations cannot keep model replacement busy until restart. (#2583, #2614, #2625, #2636)

Explicit --bits and --format pulls remain eligible for the mirror. Malformed or interrupted cache metadata no longer crashes cached-model listing, pull admission, or chat model switching. (#2610, #2613)

Explicit variant pulls persist their selected subfolder, so a later serve command resolves the pulled checkpoint instead of the repository root. Catalog aliases retain precedence over this pull marker, and mirror-backed pulls now persist the same selection as fallback downloads. (#2558, #2750)

Oversized vision images are automatically reduced to the model's patch-aware token budget before preprocessing. Multiple images share that budget, and a measured retry handles processor rounding. If the minimum aligned image remains over budget, the server warns and leaves the existing downstream prefill-cap guard as the final rejection boundary. (#2694)

Image edits derive their output canvas from the uploaded source instead of silently falling back to the text-to-image 1024×1024 default. Square and non-square FLUX.2 Klein edits now preserve the source dimensions. (#2759)

Forced assistant prefixes become visible as soon as the scheduler admits the request instead of waiting for the first decoded token. Empty, failed, and cancelled streams retire their pending admission work cleanly. (#2674)

Suffix decoding now checks the full verify-forward cache growth before advancing a sliding-window cache. At a rollback-unsafe boundary it falls back to ordinary decoding instead of aborting the request. (#2682)

A terminal MTP response that has verified farther than its visible output no longer publishes that advanced state as a reusable prefix cache. Ordinary completed-response cache reuse is unchanged. (#2751)

Parakeet v3 metadata and the Desktop picker now agree on all 25 supported ISO languages and automatic language detection instead of describing the checkpoint as English-only. (#2729)

Passive version checks correctly order rcN below the matching final and now cover pull, ps, info, bench, and doctor; automatic upgrade prompts remain off for development, RC, and local builds. (#2431)

Privacy and security

Desktop activation milestones are created and sent only after explicit telemetry consent. Request attribution uses an allowlisted client bucket; raw user-agent text never enters the payload. (#2428, #2436)

Embedded API bearer credentials support per-launch, daily, and manual rotation. The bearer stays in a code-identity-scoped Keychain item while preferences contain only non-secret metadata. (#2639)

Desktop web browsing validates every DNS answer, then pins the socket to the selected address while preserving hostname-based TLS verification. Validated IPv4 and IPv6 destinations are raced with a short stagger so one black-holed route cannot consume the whole deadline. Redirect, body-size, and timeout limits remain fail closed. (#2645, #2747)

All changes

fix(release): run promotion publishers after skipped build (7da4067)

chore: bump version to 0.13.2 (2a63362)

fix(desktop): clarify memory-limited photo mode (#2778) (00eb18a)

fix: publish promoted desktop candidates (#2775) (d7d9698)

Install: brew upgrade rapid-mlx or pip install -U rapid-mlx==0.13.2 (or just rapid-mlx upgrade).

Original source

What's new in v0.13.2-rc1

Rapid-MLX 0.13.2-rc1 makes long-running local assistants faster and more dependable: Qwen3.8 Flash-Next gains opt-in native MTP and faster long-context prefill, repeated prompts recover their prefix cache, offline speech pulls include their runtime requirements, and Desktop tightens attachment, credential, and web-browsing safety. This first release candidate is published for validation and deliberately does not replace the stable updater feed.

Highlights

Qwen3.8 Flash-Next gains native MTP — The engine can opt into the checkpoint's one-layer prediction head while the target model verifies every proposal and all recurrent, QSA, and KV state rolls back atomically. The measured fixed-K1 workload accepted 76.41% of proposals. (#2572, #2655)

Context

Serial decode

Native MTP

Change

128

25.17 tok/s

34.85 tok/s

+38.5%

2K

23.64 tok/s

33.53 tok/s

+41.8%

8K

22.82 tok/s

32.20 tok/s

+41.1%

32K

21.16 tok/s

28.82 tok/s

+36.2%

Desktop's normal temperature, top-p, top-k, min-p, and penalty settings can use the lane. Seeded requests and stateful grammar, tool, reasoning, or suppression processors continue on ordinary decoding rather than silently weakening their contract. The measured MTP path used up to 6.6 GB more active memory, and 192 GB remains the recommended hardware tier for this experimental checkpoint.

Faster long-context prefill and reusable prompts — QSA index-cache work is batched across eligible prefills, reducing measured time to first token by 28.9–32.5% while decode speed and cache precision remain unchanged. A completed 32K request can no longer poison the next request in the same process. (#2574, #2596)

Prompt

Previous TTFT

Batched TTFT

Change

2K

3.346 s

2.262 s

-32.4%

8K

13.689 s

9.236 s

-32.5%

32K

62.851 s

44.659 s

-28.9%

Semantic prefix snapshots now follow the exact rendered prompt and preserve the Flash-Next recurrent cache through batching and persistence. A measured 5,288-token warm request reused 5,273 tokens and completed in 0.539 seconds instead of 6.497 seconds; native MTP continued proposing after the hit. (#2588, #2644)

Kokoro is ready before the machine goes offline — rapid-mlx pull now fetches the voice assets and prepares the English G2P requirement as part of the pull transaction. The clean-cache validation downloaded 54 voice files and generated speech with networking disabled. If a runtime requirement is missing, an inference request returns an actionable readiness error instead of trying to download or install software on demand. (#2648, #2664)

A smaller, byte-identical-to-approved Desktop release path — LZMA packaging and dependency-proven pruning reduced the signed and notarized comparison DMG from 187,505,002 to 105,881,983 bytes, a 43.53% reduction. The tradeoff is a slower one-time Finder copy on the measured host: 17.80 seconds instead of 5.94 seconds. Candidate builds now identify their source commit, and protected publication promotes the exact approved DMG and updater payloads instead of rebuilding different bytes after the tag. Tier 1 and Desktop now install the same candidate wheel, with a shared contract matrix keeping text and multimodal request behavior aligned. (#2668, #2450, #2530, #2726)

Desktop safety and workflow corrections

After the first completed exchange, Desktop can derive one short local title without replacing a user rename. Settled text answers can also offer three optional follow-up prompts; malformed, duplicate, wrong-script, or incomplete suggestions stay hidden. (#2698)

A chat message accepts up to four images and 6 MiB of aggregate encoded image data, preserves selection order, and explains whether count or size rejected the remainder. A repeatedly failing image gets one bounded follow-up retry instead of contaminating every later text turn. (#2541, #2585)

Generated-image deletion uses an app-owned confirmation sheet: Keep remains pressable in native GUI automation, Escape is safe, and Return cannot delete an image. (#2387, #2578)

Switching away from a busy model asks first; Cancel preserves the live model and in-flight response. Recommendation, review, and engine admission use one measured footprint for known models, and already-resident models remain usable through the safe replacement path. (#2430, #2543, #2619)

The installer recommends lfm2.5-1b-4bit below 16 GB and qwen3.5-4b-4bit on larger Macs, while preferring an eligible cached model from the same memory tier. (#2426)

Exited server leaders are reaped without blocking a shared worker, and stale dual-stack tcp46 listeners are recognized when Desktop clears a server port. A stopped dictation model reloads when the user next presses the hotkey instead of during foreground activation. (#2562, #2593, #2663)

Tools settings resolve the selected search key's real Keychain state without exposing it, dictation errors no longer guess at an unproven memory cause, and photo guidance names the remedy for the engine's actual serving-lane reason. (#2514, #2523, #2602, #2607)

After 35 successful chats, dictations, or generated images, established Desktop users may see a quiet, nonmodal invitation to visit the project on GitHub. Dismissing it starts a three-day cooldown and a progressively larger local usage threshold; Rapid performs no account or star-status lookup. (#2675)

Engine, API, and CLI reliability

Qwen3.8 27B required and named tool calls use the checkpoint-native format and return OpenAI-compatible JSON arguments. Invalid required arguments fail with a client error instead of appearing executable. (#2660)

Explicit --mllm selection wins over automatic architecture, cache, and runtime fallback. The measured vision-memory floor remains mandatory, and speculative decoding still uses its supported text lane. (#2643, #2669)

Explicit timeout: 0 means the server default, request-local chat_template_kwargs reach the tokenizer, cancellations are distinct from max-length completion, and orphaned streaming reservations cannot keep model replacement busy until restart. (#2583, #2614, #2625, #2636)

Explicit --bits and --format pulls remain eligible for the mirror. Malformed or interrupted cache metadata no longer crashes cached-model listing, pull admission, or chat model switching. (#2610, #2613)

Explicit variant pulls persist their selected subfolder, so a later serve command resolves the pulled checkpoint instead of the repository root. Catalog aliases retain precedence over this pull marker, and mirror-backed pulls now persist the same selection as fallback downloads. (#2558, #2750)

Oversized vision images are automatically reduced to the model's patch-aware token budget before preprocessing. Multiple images share that budget, and a measured retry handles processor rounding. If the minimum aligned image remains over budget, the server warns and leaves the existing downstream prefill-cap guard as the final rejection boundary. (#2694)

Image edits derive their output canvas from the uploaded source instead of silently falling back to the text-to-image 1024×1024 default. Square and non-square FLUX.2 Klein edits now preserve the source dimensions. (#2759)

Forced assistant prefixes become visible as soon as the scheduler admits the request instead of waiting for the first decoded token. Empty, failed, and cancelled streams retire their pending admission work cleanly. (#2674)

Suffix decoding now checks the full verify-forward cache growth before advancing a sliding-window cache. At a rollback-unsafe boundary it falls back to ordinary decoding instead of aborting the request. (#2682)

A terminal MTP response that has verified farther than its visible output no longer publishes that advanced state as a reusable prefix cache. Ordinary completed-response cache reuse is unchanged. (#2751)

Parakeet v3 metadata and the Desktop picker now agree on all 25 supported ISO languages and automatic language detection instead of describing the checkpoint as English-only. (#2729)

Passive version checks correctly order rcN below the matching final and now cover pull, ps, info, bench, and doctor; automatic upgrade prompts remain off for development, RC, and local builds. (#2431)

Privacy and security

Desktop activation milestones are created and sent only after explicit telemetry consent. Request attribution uses an allowlisted client bucket; raw user-agent text never enters the payload. (#2428, #2436)

Embedded API bearer credentials support per-launch, daily, and manual rotation. The bearer stays in a code-identity-scoped Keychain item while preferences contain only non-secret metadata. (#2639)

Desktop web browsing validates every DNS answer, then pins the socket to the selected address while preserving hostname-based TLS verification. Validated IPv4 and IPv6 destinations are raced with a short stagger so one black-holed route cannot consume the whole deadline. Redirect, body-size, and timeout limits remain fail closed. (#2645, #2747)

All changes

fix: compare MTP ratio at exported precision (#2773) (8c47630)

fix: make MTP release probe exercise sustained decode (#2771) (15850d3)

fix: prove MTP metrics before residency sequence (#2770) (f051a6c)

fix(release): install vision runtime for top-10 gate (#2768) (dc93838)

fix(release): bind agent sequence path before cwd changes (#2766) (52d4751)

fix(release): accept Sparkle item version fields (3fa13f2)

chore: bump version to 0.13.2-rc1 (b68587a)

docs: finalize 0.13.2-rc1 release notes (#2734) (e0b6b5e)

fix: preserve image edit canvas dimensions (#2759) (3e54e4a)

test: restore Harmony terminal lifecycle fixture (#2756) (ace6b9b)

test: stub serving lane decisions in stale tests (#2719) (a18a998)

feat(chat): name conversations and suggest next questions (#2698) (36ef053)

tests: checked-in top-10 residency load-order sequence spec (#2496) (#2509) (2c47013)

test: enforce text and multimodal lane parity (#2726) (e5804f5)

fix(cli): retain mirrored pulled variant (#2750) (3016243)

fix(mtp): discard cache after terminal verify (#2751) (c0e16f3)

fix(mac): race validated DNS fallbacks (#2747) (6f90347)

fix(ci): bound cached Quickstart readiness (#2745) (aac1d98)

fix(ci): recognize trusted merge candidates (#2730) (ef4f8b0)

fix(audio): align Parakeet v3 language metadata (#2729) (b0d2a70)

fix(engine): resolve pulled --bits variant at serve/load (#2340) (#2558) (dea9284)

docs: stage 0.13.2 release notes (#2678) (ed87d72)

fix(mac): keep Quickstart starter copy truthful (#2604) (607d12a)

fix(cli): report Flash-Next native MTP capability (#2611) (fc9042c)

fix(cli): resolve the unique complete snapshot offline (#2642) (24360a8)

Fix Parakeet v3 language metadata (#2680) (0e0b927)

Fix MTP request lifecycle isolation (#2717) (b456bd8)

fix(guided): constrain Flash-Next structured output (#2620) (0346934)

fix(mac): clear bearer when selecting every start (#2718) (7061c5c)

test(gui): deliver real Settings smoke clicks (#2652) (13de792)

Prevent request-target header injection (#2705) (e4317d8)

Fix Host authority default-port handling (#2710) (4406382)

Add validated DNS address fallback (#2709) (e953ff8)

Bound chunked response parsing safely (#2704) (4efb97a)

ci(release): gate publication on real FLUX generation (#2689) (5b69b1f)

fix(mac): show per-request MTP availability (#2683) (cf54fe3)

test(mac): make lifecycle suites async and preference-isolated (#2673) (2d7c326)

fix(mac): gate residency flags by runtime support (#2688) (7448217)

Fix Host header for IPv6-pinned DNS names (#2703) (623e8da)

Fix IP-pinned transport cancellation resume (#2700) (c089775)

Downscale oversized vision inputs to token budget (#2694) (9418e03)

macOS: ask for a GitHub star after a value moment (#2675) (aae207b)

Yield forced assistant prefix after admission (#2674) (1c86ad5)

Preflight suffix decode cache growth (#2682) (7a6d40b)

test: match centralized GUI host guard (#2684) (dc19e31)

test(gui): await sidecar exit before persona cleanup (#2677) (740ce4e)

fix: honor explicit MLLM lane selection (#2669) (e9719e6)

fix(audio): make Kokoro pulls offline-ready (#2664) (6ad06cc)

fix(mac): pin browse sockets to validated IPs (#2645) (927c02e)

perf(mac): shrink the release DMG safely with LZMA (#2668) (7cc5334)

fix(mac): sweep dual-stack tcp46 listeners (#2663) (5148dc2)

fix: align Qwen3.8 tool-call wire format (#2660) (d730e18)

fix(mac): localize photo-unavailable remedies (#2607) (472d0ba)

fix(engine): apply vision memory gate to forced lane loads (#2643) (f9e49a1)

fix(mtp): serve Desktop default sampling without silent fallback (#2655) (66a0be6)

fix(cli): pull audio runtime assets for offline use (#2648) (57818d3)

test(mllm): align stats with cancelled status (#2650) (7465228)

fix(cache): restore Flash-Next prefix hits (#2644) (244849f)

fix(mac): preserve empty GUI golden arguments (#2637) (f737940)

feat(mac): persist embedded API bearer with safe lifetimes (#2639) (0b59a07)

fix(engine): release orphaned streaming admission reservations (#2636) (6747050)

test(mac): make active-switch journey fail closed (#2626) (1ae2d57)

fix(engine): distinguish cancellations from max-length completions (#2625) (55fb1dc)

test(mac): settle Launch registry before golden capture (#2630) (0857bfd)

ops: harden shared macOS host lanes (#2549) (b0623da)

ci(no-mlx): replace manual roster with requires_mlx discovery (#2534) (17d3044)

feat(telemetry): wire emit.request into /v1/messages + /v1/completions (task C, 0.13.2) (#2436) (79e2eb1)

fix(api): pass chat_template_kwargs through to apply_chat_template (#2474) (#2583) (f61a02b)

fix(residency): serve residents through reject-mode replacement loads (#2550) (#2619) (b105983)

fix(api): preserve zero request timeout as server default (#2614) (1f9db16)

fix(cli): keep mirror fallback for selected pull variants (#2554) (#2613) (0f8a6ad)

fix(cli): isolate malformed cache probe faults (#2552) (#2610) (edf5a45)

fix(mac): resolve desktop papercut states (#2602) (144b8f9)

fix(mac): defer dictation restart until the next hotkey (#2593) (38034a3)

feat(version-check): parse rcN below final; widen staleness call sites (#0.13.2 version-convergence) (#2431) (a7af663)

fix(qsa): materialize batched index keys before cache commit (#2596) (9ad8352)

fix(engine): SSOT frozensets for serving_lane_reason + contract roster (#2515) (#2544) (1c86e19)

fix(engine): preserve Flash-Next MTP cache on prefix hits (#2588) (5ddb26b)

fix(mac): reap exited server leaders without blocking a shared worker (#2363) (#2562) (78d4ad2)

fix(mac): quarantine an image after its second terminal retry failure (#2585) (a890f4c)

test(mac): table-driven memory-projection invariants across tier/picks + lane transitions (#2497) (#2563) (4c94abf)

feat(mac): report consented Desktop activation milestones (#2428) (5b17d77)

perf(qsa): batch Flash-Next compressed-key prefill (#2574) (1b9e94b)

test(gui): convert control-behaviour contract to symbol/behaviour checks (#2494) (#2548) (fa41d38)

fix(memory): use one footprint per recommended model (#2543) (b740835)

feat(engine): add native MTP for Qwen3.8 Flash-Next (#2572) (b205644)

docs: add optional contributor X-handle field to the PR template (#2532) (7ee64f5)

fix(mac): keep image deletion confirmation pressable (#2578) (49a634b)

docs: clarify queued candidate handling (#2571) (88aad52)

ci: enforce no-mac and mac queue lanes (#2575) (eee5668)

ci: split merge queue by mac requirement (#2573) (376174f)

fix(mac): bound image attachment count and aggregate bytes (#2456) (#2541) (04cee35)

ci: batch merge-ready PRs behind one integration gate (#2569) (df00905)

fix(mac): allow deleting generated images (#2387) (45c62e3)

test: declare hermetic integration inputs (#2566) (44773b6)

test(mac): bound dictation busy-waits by time, not yield count (#2537) (#2538) (ef0e752)

fix(pr_validate): reject unedited PR-template bodies in cl_description_quality (#2539) (6e0bf2c)

fix(mac): tie retained chat-lane provenance to authoritative catalog generation (#2364) (#2545) (b2216ce)

fix(ci): make parser microbench a relative regression gate (#2409) (8631acc)

ci(release): identify Desktop candidate artifacts (#2450) (8734423)

ci(l1-smoke): relocate qwen3.5-9b-4bit off hosted L1 to the Studio tier-1 gate (#2419) (#2437) (c2a64b3)

ci: reject duplicate pytest roster entries (#2531) (a8b33de)

ci: reserve expensive mac lanes for promoted heads (#2556) (c09b0e2)

test: make the hermetic HF fixture network-off by default (#2518) (#2525) (48ac10b)

ci: recover full-ci gate from cancelled duplicate runs (#2560) (2c15fab)

test(mac): prove the deferred telemetry consent boundary (#2424) (05b896e)

perf(qwen4): vectorize long-context QSA mask construction (#2533) (4219178)

feat: cached Wan checkpoints pinned by WAN_REVISIONS count as runnable (#2406 part B) (#2425) (71765d0)

fix(mac): bound hung Desktop test suite to 2 min + upload sample artifact (#2488) (#2526) (08491fc)

ci(release): promote exact pre-tag desktop artifacts (#2530) (d03cec8)

fix(installer): recommend a fast first-chat model (#2426) (d595f23)

ci(release): preflight pinned sidecar snapshots (#2529) (71693b7)

fix(mac): make the server log drawer closable (#2435) (deb10ab)

fix(mac): make serving-lane photo hints name the right remedy (#2514 follow-up) (#2523) (e2c7f6d)

fix(mac): preserve active work across model switches (#2430) (fed94eb)

docs(benchmarks): Qwen3.8-Flash-Next-4bit sampled evals vs 27B-4bit (identical harness) (#2528) (df24795)

ci: keep unpromoted full gates pending (#2517) (d7f1c53)

feat(ci): train_gates.sh reproduces the hosted gate matrix locally + drift test (#2487) (#2506) (2a9ed0f)

fix(mac): map the serving-lane reasons the engine actually emits (#2514) (83aa8b8)

docs: review cap back to 20 rounds; reviewers must not expand PR scope (#2516) (4181499)

fix(mac): align onboarding admission with the curated-pick predicate so a recommended model is never a dead-end (#2505) (#2512) (e91b2da)

ci(release): required bump-PR guard + dispatch-bound release preflight; fix #2491 (#2511) (436a7f6)

tests: hermetic HF cache/env fixtures + declared lane dep extras; fix #2445/#2446 (#2508) (c8ee8ce)

feat(pr): contract-field template, pr_validate exact merge-base, body-only precheck (#2507) (205ac98)

fix: keep vision extra on the validated runtime (#2501) (3e6afa1)

docs: publish Qwen3.8 Flash-Next M3 Ultra evidence (#2504) (c9fc2e6)

test(mac): harden desktop tests for 0.13.2 (#2480 RAM-independent re-enable + #2481 flaky fix) (#2500) (9456f54)

feat(cli): cheetah launch banner (interactive only) (#2485) (567e9dd)

ci: add auto-release dry run (#2503) (5b6841c)

Community contributors

@osdodo — fix(mac): allow deleting generated images

@guo — fix(mac): make the server log drawer closable

@Jevin-F — fix(mac): map the serving-lane reasons the engine actually emits

@guo — Fix Parakeet v3 language metadata

Install: brew upgrade rapid-mlx or pip install -U rapid-mlx==0.13.2-rc1 (or just rapid-mlx upgrade).

Original source

[0.13.2] — 2026-08-30

Rapid-MLX 0.13.2 makes long-running local assistants faster and more reliable, improves offline speech and Desktop safety, and promotes the exact signed Desktop candidate that passed release validation.

Added

Opt-in native MTP for Qwen3.8 Flash-Next. Target verification and atomic recurrent-state rollback raise measured decode throughput by 36–42% across 128-token through 32K-context workloads while constrained requests retain ordinary decoding.

Conversation titles and follow-up suggestions. Completed local chats can derive a short title and offer three optional next steps without replacing a user rename or displaying malformed output.

Privacy-bounded activation milestones. Desktop records the first successful chat, dictation, and generated image only after explicit consent.

Changed

Flash-Next long-context prefill is faster and reusable. Batched QSA index-cache construction reduced measured 2K, 8K, and 32K time to first token by 28.9–32.5%, and semantic snapshots preserve reusable recurrent state through batching and persistence.

The Desktop download is smaller. LZMA packaging and dependency-proven pruning reduced the signed comparison DMG by 43.53% while retaining the release contract.

First-chat recommendations follow available memory. New installs choose a smaller default below 16 GB and prefer an eligible cached model in the same memory tier.

Fixed

Photo limits no longer make text chat look broken. When a model's vision lane needs more memory than the Mac has, Desktop now says that text chat is still ready, recommends a lower-memory vision model, and dismisses the notice after the user continues. (#2778)

Offline Kokoro pulls include their runtime assets, image attachments enforce count and encoded-size budgets, and generated-image deletion uses a stable app-owned confirmation sheet.

Model switching, request cancellation, required tool calls, explicit model variants, oversized vision inputs, image-edit dimensions, suffix rollback, MTP cache publication, and multilingual Parakeet metadata now retain their validated contracts across the engine and Desktop.

Web browsing pins validated IPv4 and IPv6 destinations while preserving TLS hostname verification, and Desktop credentials stay in Keychain-backed, rotation-aware storage.

Protected publication promotes the exact signed and notarized Desktop candidate bytes that passed pre-tag validation. (#2775)

Original source

[0.13.2-rc1] — 2026-08-30

Added

Opt-in native MTP for Qwen3.8 Flash-Next. The engine can use the checkpoint's one-layer prediction head with target verification and atomic rollback. On the measured 128-token through 32K-context workloads, decode rose from 21.16–25.17 tok/s to 28.82–34.85 tok/s with 76.41% proposal acceptance. Desktop's normal temperature and penalty settings are supported; seeded or stateful constrained requests continue on ordinary decoding.
(#2572,
#2655)

Consented Desktop activation milestones. Rapid can report the first successful chat reply, dictation, and generated image after explicit opt-in. Successful Messages and Completions requests also carry privacy-bounded surface and client attribution; raw user-agent text is never emitted.
(#2428,
#2436)

Conversation titles and next-step suggestions. After the first completed exchange, Desktop can derive one short local title without overwriting a user rename. Settled text answers can also offer three optional follow-up prompts; malformed, duplicate, or incomplete suggestions remain hidden.
(#2698)

Changed

Long-context Flash-Next prefills are faster. Batched QSA index-cache construction reduced measured 2K, 8K, and 32K time to first token by 28.9–32.5% without changing decode speed or cache precision. A follow-up evaluation boundary keeps a completed 32K request from corrupting the next request in the same process. (#2574,
#2596)

Repeated Flash-Next prompts reuse their prefix. Semantic snapshot boundaries now follow the exact rendered request, including request-local template options, and preserve the model-specific recurrent cache through batching and persistence. In the measured 5,288-token prompt, a warm request reused 5,273 tokens and completed in 0.539 seconds instead of 6.497 seconds; native MTP remained active after the hit.
(#2588,
#2644)

The installer recommends a faster first-chat model that fits the Mac. Fresh installs below 16 GB suggest lfm2.5-1b-4bit; larger Macs suggest qwen3.5-4b-4bit, while an eligible cached curated model remains preferred.
(#2426)

Release-candidate builds can see final-version updates. Passive version notices order rcN below the matching final release and now appear in pull, ps, info, bench, and doctor. Automatic upgrade prompts remain disabled for development, RC, and local builds.
(#2431)

The Desktop download is substantially smaller. Release packaging uses LZMA and removes only build-time or dependency-proven-unused sidecar files. The signed and notarized comparison artifact fell from 187,505,002 bytes to 105,881,983 bytes (43.53% smaller); the measured Finder copy became slower, from 5.94 seconds to 17.80 seconds.
(#2668)

A GitHub invitation waits for demonstrated value. After 35 successful chats, dictations, or generated images, established Desktop users may see a quiet, nonmodal project invitation. Dismissal starts a three-day cooldown and a progressively larger local workload threshold; no account or star-status lookup is performed. (#2675)

Fixed

Kokoro pulls are complete before the machine goes offline. rapid-mlx pull now fetches the voice assets and prepares the English G2P requirement alongside the checkpoint. Inference never downloads or installs a missing runtime component from inside a request; it returns an actionable readiness error instead. (#2648,
#2664)

Parakeet v3 reports its complete language support. Registry metadata and the Desktop picker now agree on all 25 supported ISO languages and automatic language detection instead of describing the checkpoint as English-only.
(#2729)

Qwen3.8 tool calls use the checkpoint's native wire format. Required and named tool calls return OpenAI-compatible JSON arguments, and malformed required arguments fail with a client error instead of looking executable.
(#2660)

Explicit multimodal selection stays explicit. --mllm takes precedence over automatic architecture, cache, and runtime fallbacks. The measured vision-memory floor still fails closed, and speculative decoding continues to select its supported text lane.
(#2643,
#2669)

Photo attachments have bounded, recoverable behavior. A message accepts at most four images and 6 MiB of aggregate encoded image data, reports which limit rejected the remainder, and cannot resend one permanently failing image forever. Generated-image deletion uses an app-owned confirmation sheet whose safe Keep action remains pressable and whose Return key cannot delete.
(#2541,
#2585,
#2387,
#2578)

Model changes preserve active Desktop work. Switching away from a busy model asks before replacement, Cancel keeps the response alive, and the shared measured footprint keeps recommendation, review, and engine admission aligned. Already-resident models remain servable through the safe replacement path. (#2430,
#2543,
#2619)

Desktop lifecycle recovery no longer blocks shared workers. Exited server leaders are reaped after the kernel exit event, dual-stack tcp46 listeners are recognized when clearing a stale server port, and a stopped dictation model reloads on the next hotkey instead of during foreground activation.
(#2562,
#2593,
#2663)

Settings and photo guidance describe the actual state. Tools shows whether the selected search key is present without exposing it; dictation no longer guesses that every preparation failure is a memory problem; photo remedies use the engine's real serving-lane reasons and point to the correct action. (#2514,
#2523,
#2602,
#2607)

OpenAI-compatible request behavior is more predictable. Explicit timeout: 0 means the server default, request-local chat_template_kwargs reach the tokenizer, cancellations are no longer mislabeled as max-length completion, and orphaned streaming reservations cannot leave model changes busy until restart. (#2583,
#2614,
#2625,
#2636)

Pull and cache recovery fail locally without losing the fast path. An explicit --bits or --format selection remains eligible for the mirror, and malformed or interrupted cache metadata no longer crashes cached-model listing, pull admission, or chat model switching.
(#2610,
#2613)

Pulled variants resolve consistently on serve. Pulling an explicit --bits or --format variant records the selected subfolder so serve loads the same checkpoint instead of the repository root. Catalog aliases retain precedence over the pull marker, and mirror-backed pulls now commit the same selection as fallback downloads.
(#2558,
#2750)

Oversized vision inputs fit the request budget automatically. Images are reduced against the model's patch-aware token ceiling before preprocessing, with one measured retry for processor rounding. If the minimum aligned image remains over budget, the server warns and leaves the existing downstream prefill-cap guard as the final rejection boundary.
(#2694)

Image edits preserve the source canvas. Editing an uploaded image derives the output dimensions from that image instead of silently selecting the text-to-image 1024×1024 default. Square and non-square FLUX.2 Klein edits retain their original dimensions. (#2759)

Forced assistant prefixes no longer wait for the first decoded token. The prefix is emitted after scheduler admission, while empty, failed, and cancelled streams clean up their pending admission work.
(#2674)

Suffix decoding respects sliding-window rollback boundaries. The engine preflights the full verify-forward cache growth and falls back to ordinary decoding when advancing would make rollback unsafe, rather than aborting the request. (#2682)

Terminal MTP responses cannot publish speculative cache state. When a response finishes before every verified draft token is emitted, Rapid drops that response's reusable cache rather than exposing state ahead of the visible output. (#2751)

Security

Embedded API credentials have explicit safe lifetimes. Per-launch, daily, and manual-rotation policies store the bearer in a code-identity-scoped Keychain item; preferences contain only non-secret metadata. Unavailable or malformed persisted credentials degrade to a one-time key and are surfaced in Settings. (#2639)

Desktop web browsing pins the validated destination. Every DNS answer is checked by the existing SSRF policy, then the socket connects directly to the selected address while retaining the original hostname for TLS and certificate validation. Validated IPv4 and IPv6 destinations are raced with a short stagger, so one black-holed route cannot consume the whole browse deadline. This closes the DNS-rebinding window without weakening redirect, body-size, or timeout limits.
(#2645,
#2747)

Release engineering

Signed Desktop candidates identify their exact source commit, and protected publication promotes the already reviewed DMG, Sparkle archive, appcast, and release notes instead of rebuilding different bytes after the tag. The same candidate wheel is installed in Tier 1 and promoted into the Desktop sidecar, while text and multimodal lanes share an explicit request-contract matrix.
(#2450,
#2530,
#2726)

Original source

What's new in v0.13.1

Rapid-MLX 0.13.1 focuses on safer model changes and smoother first-use paths:

assistant replacement is now an explicit transaction, large multi-variant repositories can download only the requested format, and mirrored models stay on the fast download route.

Highlights

Experimental Qwen3.8-Flash-Next text inference

The M1 lane supports the mixed 4-bit checkpoint through the qwen3.8-flash-next-4bit alias on Macs with at least 128 GB of unified memory. This release exposes the experimental text lane through the CLI and API; release validation used a 256 GB Mac, and Desktop, MTP, and vision support remain planned for a later release.

(#2433)

Switch assistants without losing track of active work

API model replacement now has explicit reject, wait, and abort policies shared by the engine and server; Desktop uses the safe reject policy while a request is busy. Request ownership transfers atomically, admitted work is quiesced before reload, and auxiliary dictation residency remains independent of the selected assistant.

The /v1/models/load API defaults memory_policy to evict_first_if_needed: it keeps the rollback-safe load when old and new fit together, evicts the replaced assistant first only when keeping both breaches the limit but the replacement alone fits, and otherwise returns a typed 507 with replacement_projection before destructive mutation. Set memory_policy=keep_then_commit to require rollback-safe loading.

(#2369)

Download one usable variant instead of an entire repository

rapid-mlx pull accepts --bits or --format for repositories that store several quantizations side by side. This avoids turning a small-model download into tens of gigabytes of formats the user did not select.

(#2145, #2338)

More dependable downloads and tool routing

Subfolder-based quantizations can use the Rapid mirror instead of being hard-routed upstream. Dedicated evaluation scenarios now catch explicit current-weather prompts that choose a generic search fallback instead of the advertised Weather tool.

(#2279, #2222, #2327)

Photo input uses the validated bundled vision runtime

The Desktop sidecar carries the updated vision stack and the release gate now proves that two cached vision families produce image-dependent answers before an artifact can ship.

(#2384, #2380)

Vision-capable Qwen3.5, Qwen3.6, and supported Qwen3.8 checkpoints use the vision lane automatically when mlx-vlm 0.6.16 or newer is installed and the checkpoint meets its measured vision-memory recommendation. If automatic admission cannot use that lane, serving falls back to text with a machine-readable reason. The experimental Flash-Next checkpoint is explicitly text-only in this release. CLI users can force text serving with --no-mllm, and Desktop exposes the same per-model Performance choice. Requests that enable MTP or another speculative decoder select the text lane, where that decoder is supported.

A better first model for each Mac

First run now prefers an eligible cached model, then chooses a starter that fits the machine instead of applying one recommendation to every Mac. Supported hybrid vision checkpoints are routed to their working multimodal lane, so photos work with the recommended model without a manual engine choice.

(#2385, #2219, #2388)

Reliability and setup corrections

API request validation accepts either a scalar or array stop, rejects invalid timeouts, requires real JSON booleans for resident-model controls, and reports the precise invalid field path.

(#2367, #2371, #2372)

Guided setup now says explicitly that it preserves user data, avoiding a misleading destructive-action warning.

(#2239, #2322)

The one-time telemetry consent question waits until Rapid has delivered a successful chat reply, dictation transcript, or generated image instead of interrupting first launch. Nothing is sent before explicit consent.

(#2424)

Share CLI coverage now exercises only supported invocation forms and runs in CI.

(#2377, #2382)

rapid-mlx doctor distinguishes the executable running inside the active environment from another installation found on PATH, and explains how to resolve a mismatch.

(#2352, #2402)

Serving an uncached model while offline stops with one actionable cache and connectivity explanation instead of retrying the same download or suggesting an unrelated serving lane.

(#2357, #2423)

8 GB Macs start with a model that fits. First run recommends lfm2.5-1b-4bit on the lowest memory tier, avoiding the RAM warning produced by the previous 2.6B starter.

(#2432)

Dictation stays resident and is restored after switching chat models.

(#2400)

MTP generation remains bound to its worker when another model loads, so the original assistant continues answering instead of losing its GPU stream.

(#2441)

Model-switch memory estimates credit the assistant being replaced and use the same fit thresholds as the actual replacement transaction. Desktop's low-memory threshold is now 100% instead of 85%: projections from 95–100% are advisory, while values above 100% require confirmation. The earlier conservative threshold is superseded because the projection now credits the model being replaced, avoiding a warning for memory that will be freed before the new model starts.

(#2443, #2444)

Speech to Text automatically arms after model lifecycle changes when the user has enabled it; there is no extra manual start button to press.

(#2448)

JSON-schema responses and bounded-thinking or title-generation requests now terminate normally on the vision lane: their request-local output processors and complete configured EOS set reach the multimodal scheduler, while photo requests continue using the same server.

(#2471)

Experimental Qwen3.8-Flash-Next remains on its supported text lane; MTP and other speculative decoding requests also choose the text lane; and text-diffusion assistants can replace another assistant without a spurious group-conflict response.

(#2472)

iPhone HEIC photos work through picker, drop, and paste. Supported still images are normalized to truthful JPEG or PNG bytes at the shared attachment boundary while preserving the attachment size limit.

(#2467)

High-resolution photos are downscaled before Desktop sends them so they fit the model's vision budget, and a typed image rejection no longer leaves a failed attachment turn that contaminates the conversation.

Launch no longer rehashes every sealed app resource merely to choose the Keychain namespace, and Escape cannot silently decline the post-value telemetry-consent invitation.

(#2470)

A rejected photo no longer contaminates the next text turn, and multimodal unload/reload cannot reuse a stale generation worker.

(#2379, #2378, #2401, #2397)

Readiness and connection guidance now report the live port, selected model, and served name. Public streaming request IDs can be cancelled reliably.

(#2386, #2348, #2395, #2353, #2398, #2342)

Failed primary reload recovery keeps registry, residency, routing, readiness, and speech handoff consistent.

(#2394, #2360)

Downloads retain the catalog's size warning when remote metadata is missing, and cached pulls are reported as verified rather than newly downloaded.

(#2391, #2350, #2392, #2349)

Complete cached Kokoro and Whisper Turbo snapshots are recognized as runnable using their verified audio weight layouts.

(#2406, #2417)

System-prompt settings distinguish the saved instruction from the effective prompt and avoid touching optional credentials until they are needed.

(#2341)

A complete offline cache with one immutable snapshot remains runnable even when a branch ref is unavailable; ambiguous multi-snapshot caches remain unresolved rather than guessing.

(#2351, #2404)

Server discovery for model switching uses live socket ownership instead of scanning mounted filesystems, avoiding unrelated volume-access prompts.

(#2343, #2408)

Native file URL drops on the chat composer use the existing attachment importer, and unsupported drops cannot insert a local path into the message.

Contributed by osdodo.

(#2396)

Release evidence and corrections

Corrected the published Nemotron Labs Diffusion description so it accurately describes the supported text-model path.

(#2376)

Recorded a reproducible M2 Pro performance comparison between the published 0.13.0 and 0.12.18 packages.

(#2375)

Known issues

Suffix decoding with sliding-window models such as Gemma 4 and GPT-OSS can abort a request at a window boundary. Disable suffix decoding for these models as a workaround.

(#2463)

All changes

chore: bump version to 0.13.1 (#2482) (819db66)
docs: stage 0.13.1 release notes (#2393) (3e0bc30)
release: integrate 0.13.1 train 6d (#2479) (c730e45)
release: integrate 0.13.1 train 6c (f911af1)
release: integrate 0.13.1 train 6b (2063517)
release: integrate 0.13.1 train 6 (#2449) (aaf2e01)
ci: union Apple MLX coverage into changed-lines gate (#2440) (a6b820c)
fix(mac): use the safe 8 GB Quickstart model (#2432) (aaa2b8a)
release: integrate 0.13.1 train 5 (62a038c)
release: integrate 0.13.1 train 4 (19616c2)
release: integrate 0.13.1 train 3 (32288e3)
release: integrate 0.13.1 train 2 (eb9b335)
release: integrate 0.13.1 train 1 (854d049)
test(lifecycle): cover cancelled sibling retirement (36265f5)
fix(lifecycle): propagate post-commit cancellation (cb837dd)
test(residency): cover grouped reload transaction (286acfc)
fix(residency): quiesce admitted requests before reload (84bb0f1)
test(lifecycle): assert committed stop-failure routing (8b636bb)
fix(lifecycle): commit handoff before engine retirement (7034584)
fix(residency): validate explicit replacement groups (838855e)
test(lifecycle): close exact-head validation gaps (361db6b)
fix(server): materialize assistant before lifecycle abort (9ccc6d4)
fix(server): preserve lifecycle abort reason in SSE (d8cc848)
test(server): pin dictation across assistant transitions (28da6b7)
fix(lifecycle): preserve precommit abort ownership (ae1a05c)
fix(lifecycle): close replacement convergence gaps (6ab3eac)
fix(server): wait for resident leases before replacement (7c935dc)
fix(server): translate route-bound lifecycle aborts (bcd11ce)
fix(server): reject busy replacement before eviction (3dfd886)
fix(engine): transfer non-stream admission on commit (d2daf07)
fix(engine): publish scheduler requests atomically (ee3f184)
fix(server): surface pre-scheduler replacement aborts (4e30984)
fix(engine): abort pre-scheduler admissions (527a2a6)
fix(engine): preserve concurrent admission ownership (5d0d9b6)
test(server): lock dictation coexistence during replacement (d8552b7)
test(server): satisfy lifecycle type gate (073009a)
feat(server): add explicit assistant replacement policy (7470607)
docs(perf): compare v0.13.0 with v0.12.18 on M2 Pro (5f8984d)
feat(pull): select one variant of a multi-variant repo with --bits/--format (#2145) (#2338) (43afedb)
docs: correct Nemotron 0.13.0 wording (df31169)
test(evals): lock explicit-weather routing to the weather tool (#2222) (#2327) (30be6a1)
test(mirror): cover subfolder filtering in Linux lane (1c81f80)
fix(mirror): serve subfolder-per-quant repos from R2 instead of hard-declining (bc45286)

Install: brew upgrade rapid-mlx or pip install -U rapid-mlx==0.13.1 (or just rapid-mlx upgrade).

Original source

[0.13.1] — 2026-08-26

Added

Experimental Qwen3.8-Flash-Next text inference. The M1 lane supports the mixed 4-bit checkpoint through the qwen3.8-flash-next-4bit alias on Macs with at least 128 GB of unified memory. This release exposes the experimental text lane through the CLI and API; release validation used a 256 GB Mac, and Desktop, MTP, and vision support remain planned for a later release.
(#2433)

Assistant changes are explicit, safe transactions. API clients can choose whether a busy text or vision model switch should reject, wait, or abort active assistant work. Desktop rejects a busy switch safely, and auxiliary speech models remain resident. The /v1/models/load API defaults memory_policy to evict_first_if_needed: it keeps the rollback-safe load when old and new fit together, evicts the replaced assistant first only when keeping both breaches the limit but the replacement alone fits, and otherwise returns a typed 507 with replacement_projection before destructive mutation. Set memory_policy=keep_then_commit to require rollback-safe loading.
(#2369)

Large multi-variant repositories can download one serving format. The CLI accepts --bits or --format, so users do not need to fetch every quantization in a repository. (#2145,
#2338)

Fixed

Mirrored subfolder quantizations download from the fast path. Models such as LFM2.5 2.6B 4-bit no longer fall back to a slow or unreachable upstream route merely because the quantization lives in a repository subfolder.
(#2279)

Explicit weather requests have dedicated regression coverage. Evaluation scenarios detect models selecting general web search or claiming the advertised Weather tool is unavailable. (#2222,
#2327)

Desktop photo input uses the updated bundled vision runtime. Release validation now requires image-dependent answers from two cached vision families before the sidecar can ship. (#2384,
#2380)

Vision-capable Qwen checkpoints choose the working lane automatically. Vision-capable Qwen3.5, Qwen3.6, and supported Qwen3.8 checkpoints use the vision lane automatically when mlx-vlm 0.6.16 or newer is installed and the checkpoint meets its measured vision-memory recommendation. If automatic admission cannot use that lane, serving falls back to text with a machine-readable reason. The experimental Flash-Next checkpoint is explicitly text-only in this release. CLI users can force text serving with --no-mllm, and Desktop exposes the same per-model Performance choice. Requests that enable MTP or another speculative decoder select the text lane, where that decoder is supported.

API validation fails early with actionable fields. Requests accept scalar or array stop, reject invalid timeouts and non-boolean residency controls, and identify the exact invalid load field. (#2367,
#2371,
#2372)

Guided setup accurately describes data preservation. Its confirmation no longer implies that onboarding reset deletes user data. (#2239,
#2322)

Telemetry consent appears after Rapid demonstrates value. The one-time question follows a successful chat reply, delivered dictation transcript, or generated image instead of interrupting first launch. Nothing is sent before explicit consent. (#2424)

Share CLI contracts use supported invocation forms and run in CI.
(#2377,
#2382)

Environment diagnosis identifies the CLI that is actually running. rapid-mlx doctor reports both the active-environment executable and a different installation found on PATH, with actionable mismatch guidance.
(#2352,
#2402)

Offline serving failures are concise and actionable. An uncached model now stops after one cache and connectivity explanation instead of repeating download work or recommending an unrelated serving lane.
(#2357,
#2423)

First run chooses a hardware-fit starter and photos use the correct lane. Eligible cached models remain preferred, and supported hybrid vision models route automatically. (#2385,
#2219,
#2388)

8 GB Macs start with a model that fits. First run recommends lfm2.5-1b-4bit on the lowest memory tier, avoiding the RAM warning produced by the previous 2.6B starter. (#2432)

Dictation remains available across chat-model switches. The speech model stays resident and is restored after switching chat models.
(#2400)

MTP assistants survive loading a second model. Generation remains bound to its worker, so the original assistant continues answering instead of losing its GPU stream. (#2441)

Model-switch memory projections credit the assistant being replaced. The estimate uses the same fit thresholds as the replacement transaction. Desktop's low-memory threshold moves from 85% to 100%: 95–100% shows advisory guidance and above 100% requires confirmation. The previous conservative threshold is superseded now that projections credit memory freed by the model being replaced.
(#2443,
#2444)

Speech to Text arms itself after model lifecycle changes. When speech input is enabled, it becomes ready without an extra manual start button.
(#2448)

Structured text requests terminate normally on the vision lane. JSON-schema responses and bounded-thinking or title-generation requests now terminate normally on the vision lane: their request-local output processors and complete configured EOS set reach the multimodal scheduler, while photo requests continue using the same server.
(#2471)

Serving lanes honor the requested workload. Experimental Qwen3.8-Flash-Next remains on its supported text lane, MTP and other speculative decoding requests choose the text lane, and text-diffusion assistants can replace another assistant without a spurious group-conflict response. (#2472)

iPhone HEIC photos can be attached to chat. Picker, drop, and paste share one normalization boundary that converts supported still images to truthful JPEG or PNG bytes while preserving the attachment size limit.
(#2467)

High-resolution photos fit the model's vision budget. Desktop downscales oversized images before sending them, and a typed image rejection no longer leaves a failed attachment turn that contaminates the conversation. Launch avoids full app-bundle identity revalidation. Keychain namespace selection checks the signing certificate without rehashing every sealed resource, and Escape no longer silently declines the later telemetry-consent invitation. (#2470)

Photo rejection and multimodal reloads recover cleanly. A rejected image cannot poison the next text turn, and a reloaded vision model cannot reuse a stale generation worker. (#2379,
#2378,
#2401,
#2397)

Live API identity and cancellation are dependable. Readiness and connect guidance use the live port and served model name, and public streaming IDs are cancellable. (#2386,
#2348,
#2395,
#2353,
#2398,
#2342)

Reload failures preserve a truthful serving state. Registry, residency, routing, readiness, and speech handoff remain consistent even when both a primary reload and its restoration fail. (#2394,
#2360)

Download status and size warnings stay truthful offline. Catalog size is retained when remote metadata is unavailable, and cached pulls report verification rather than a new download. (#2391,
#2350,
#2392,
#2349)

Cached speech checkpoints are recognized by their real layouts. Complete Kokoro and Whisper Turbo snapshots count as runnable when their verified family-specific weights are present. (#2406,
#2417)

System-prompt settings explain what is actually sent. Saved and effective prompts are distinguished, and optional credentials are read only when needed. (#2341)

Complete offline snapshots remain available without a branch ref. A single immutable cached revision is recognized as runnable, while ambiguous multi-snapshot caches still fail closed. (#2351,
#2404)

Model switching no longer scans mounted filesystems to discover servers. Live socket ownership supplies the same port and process facts without prompting for access to unrelated volumes. (#2343,
#2408)

Files can be dropped directly into the chat composer. Supported native file URLs use the existing attachment importer, while unsupported drops cannot leak a local path into the message. Contributed by osdodo. (#2396)

Documentation

Corrected the 0.13.0 Nemotron Labs Diffusion description to identify its supported text-model path. (#2376)

Published a reproducible M2 Pro comparison of 0.13.0 and 0.12.18. (#2375)

Known issues

Suffix decoding with sliding-window models such as Gemma 4 and GPT-OSS can abort a request at a window boundary. Disable suffix decoding for these models as a workaround. (#2463)

Original source

What's new in v0.13.0

Rapid-MLX 0.13.0 makes first setup clearer, expands local model support, keeps chat and speech workloads available together, and improves long-prompt responsiveness and cache correctness.

Highlights

More capable local models — Run supported Qwen3-Next checkpoints larger than available memory with disk streaming, serve Nemotron-Labs-Diffusion 3B as a standard text model, and choose verified Ornith 1.5 9B and 35B-A3B models from the Desktop catalog. (#2192, #2194, #2197)

Faster, safer long conversations — Hybrid models reuse compatible prefix caches, cutting verified repeat-turn prefill from tens of seconds to sub-second latency. MLX 0.32.1 and measured per-model prefill defaults also improve throughput while preserving explicit server overrides. (#2321, #2199, #2210)

Runtime model changes keep the same identity — Models loaded after startup use the same serving-lane resolution as the initial model, and performance reloads preserve the names and aliases clients already use. (#2339)

Clearer setup and model switching — First setup uses live RAM-based recommendations, groups cached variants, and reports download progress. Chat receives authoritative local date and time, active-request model switches ask before interrupting work, and complete cached Qwen3.5 checkpoints route through the correct runtime automatically. (#2331, #2334, #2335, #2336)

Chat, vision, audio, and images work together more naturally — Desktop adds a dedicated photo flow, separates Speech to Text from Text to Speech, and keeps ordinary dictation from evicting the active conversation model. Relaunch also restores chat before the enabled dictation lane, so speech cannot replace the selected conversation model. Image generation starts at a lower-memory 512×512 default and shows a stable remaining-time estimate. (#2201, #2188, #2307, #2302, #2339)

More reliable tools and reasoning — Tool selection is schema-driven, malformed tool attempts stay inside the correction loop, reasoning scratch work stays out of final answers, and supported web-search providers recover from rejected saved keys without dropping the original request. (#2244, #2278, #2284, #2304)

Upgrade notes

From 0.13.0-rc2, quit the existing app, replace it in Applications with the final DMG, and reopen it. Conversations, downloaded models, and preferences remain in place.

Existing cached dictation selections are preserved. Incompatible old prefix caches remain on disk but are not reused.

Standalone servers opt into the shared audio lane with --enable-audio; Desktop enables it automatically.

Known issues

If both a resident model's performance reload and its rollback fail, residency status and request routing can disagree until the server is relaunched. This recovery edge case is planned for 0.13.1. (#2360)

Loading a secondary model can leave the reported resident primary model returning errors until Rapid-MLX Desktop is relaunched. (#2333)

Run setup again currently overstates what it resets; conversations, downloaded models, preferences, and telemetry consent are preserved. (#2239)

Low-memory coexistence does not yet provide complete cross-role capacity guidance for every chat, vision, speech-input, and speech-output combination. (#2305)

All changes

docs: add 0.13.0 changelog (#2368) (4f9f916)

chore: bump version to 0.13.0 (2890fab)

release: integrate final 0.13.0 fixes (#2337) (a8f507a)

fix(mac): onboarding 'Before loading' memory verdict re-evaluates live (ONBOARD-MEM-LIVE) (#2331) (240a6d7)

test(ci): cover hybrid cache admission on Linux (22b999f)

fix(server): keep hybrid entry type explicit (4fd032b)

fix(cache): align exact recurrent prefix reuse (d4b81e8)

test(cli): isolate port preflight after rebase (b74820e)

fix(config): preserve strict detection failures (c726840)

fix(server): align standalone hybrid cache admission (f3bea45)

test(cli): isolate serve metadata wiring (eccad9e)

fix(cli): preserve explicit startup overrides (5e62429)

fix(config): preserve checkpoint architecture precedence (6f5ef99)

fix(cache): honor detected hybrid metadata (2098d58)

fix: preserve release notes artifact path (a45feb8)

fix(mac): make bad-descriptor drain deterministic under parallel suite (#2318) (#2323) (5bef1ea)

chore: bump version to 0.13.0-rc2 (33dd7a3)

fix(release): gate reruns before updater mutation (33cf880)

fix(release): no-op identical tagged reruns (a6c3b8c)

test(release): follow shared desktop cap contract (7b0e961)

fix(release): align tagged publication lifecycle (ea23047)

fix(release): make desktop publication immutable (4e89fde)

fix(release): compare DMG growth in bytes (dfc2a19)

fix(release): bind desktop publish to run artifact (417f60b)

fix(release): fail closed on DMG baseline lookup (e449043)

fix(release): wait for Desktop publication before engine release (109481c)

fix(release): require admin bypass disabled + honest cutoff-guard docs (#2301) (b869d3d)

feat(release): add normal workflow_dispatch retry_version after main drift (#2301) (672f091)

style(release): satisfy ruff check/format for new release contracts (#2301) (dbd95df)

fix(release): claim Desktop RC tag only at a validated commit (#2301) (eae5875)

test(mac): settle approval observers once (bf2ff33)

test(mac): synchronize declined-tool approval observation (b4e3c3f)

test(desktop): cover first-poll lifecycle contract (cb317ec)

test(mac): synchronize HF cache monitor polling (e82e3aa)

fix(mac): invalidate stale memory refresh on activation (8c8a9d9)

fix(mac): stop hidden memory warning refreshes (f6cd8d3)

fix(mac): preserve memory confirmation through alert dismissal (f639006)

fix(mac): recheck memory before resolving load (bd539fa)

fix(mac): resolve latest memory warning facts (4389635)

fix(mac): cancel foreground memory probes (7193f5f)

refactor(mac): resolve memory decision before launch (d0b6c86)

fix(mac): bind foreground refresh to onboarding (3000a03)

fix(mac): cancel parked memory refreshes (88d4c76)

test(mac): pin blocking memory verdict (759ff20)

fix(mac): order live memory refreshes (cfca404)

fix(mac): treat tight memory as normal load (c0f9e5f)

fix(mac): refresh onboarding memory safety live (d40e607)

test(mac): make startDownload re-resolve reliable under parallel suite (#2237) (#2312) (f984de2)

fix(desktop): keep conversation model resident during dictation (#2307) (21dee04)

test: bind image recovery alias to registry capability (ac471e3)

fix: recommend valid image model alias (226e6e0)

fix(mac): stabilize image generation ETA (#2302) (16bccaa)

feat(server): own audio lifecycle on the model worker (#2282) (1629a51)

feat(mac): make Connect your agents useful before an engine starts (#2308) (896e3ab)

fix(mac): recover rejected web-search keys (#2304) (cbaa801)

fix(mac): preserve dictation checkpoint across rc1 upgrade (#2303) (dee72f2)

ci(mac): raise raw bundle cap to 550 MB (b42b029)

test(mac): cover attachment retry and relaunch [skip-version-bump] (#2289) (ad79fd4)

Install: brew upgrade rapid-mlx or pip install -U rapid-mlx==0.13.0 (or just rapid-mlx upgrade).

Original source

[0.13.0] — 2026-08-26

Rapid-MLX 0.13.0 makes first setup clearer, expands local model support, keeps chat and speech workloads available together, and improves long-prompt responsiveness and cache correctness.

Added

More capable local models. Supported Qwen3-Next checkpoints can stream experts from disk, Nemotron Labs Diffusion 3B runs as a standard text model, and verified Ornith 1.5 models are available from the Desktop catalog.

Chat understands the current local date and time. Questions about today no longer depend on the model's training cutoff or require a web search.

Desktop workflows cover more of the local stack. Attachments, dedicated photo input, separate Speech to Text and Text to Speech controls, and clearer agent setup make multimodal and tool-assisted work easier to configure.

Changed

Repeated long prompts start much faster on hybrid models. Compatible prefix caches are reused, reducing verified repeat-turn prefill from tens of seconds to sub-second latency.

Setup guidance follows current memory conditions. First-run recommendations are re-evaluated as available memory changes instead of retaining a stale verdict.

Active work is protected during model switches. Desktop asks for confirmation before a switch interrupts requests already in progress.

Fixed

Runtime model switches now choose the same serving lane as startup, including complete cached checkpoints that need the text-only lane.

Performance reloads preserve the model's served names and aliases, so clients can continue using the same identifier after settings change.

Relaunching with dictation enabled restores the conversation model before the speech lane, preventing an audio model from replacing the active chat model.

Search-provider key failures, model readiness progress, image-generation recovery, tool-call correction, reasoning privacy, and release validation received the fixes exercised across the 0.13 release candidates.

Known issues

If both a resident model's performance reload and its rollback fail, residency status and request routing can disagree until the server is relaunched. This recovery edge case is planned for 0.13.1 (#2360).

Original source