Ollama Release Notes

What's Changed

Claude Desktop

Ollama now integrates with Claude

Turn individual Ollama models on or off for use in Claude, directly from the menu bar

Choose from your available Ollama models from within Claude; cloud models appear only when you're signed in

A new Apps view manages app integrations with copyable commands

Improved caching

Fixed a hang where agent clients that cancel long prefills

Prefill restore points are now trustworthy by construction: a cancelled prefill keeps every restore point it crossed, so retries resume where they stopped instead of restarting from scratch

Resumed prefills no longer record restore points that fail to cover what they claim; on models with recurrent layers this previously forced a request matching 46k of 47k tokens to reprocess from zero

Disabled Claude Code's "tokens left" token-countdown system message, which Ollama moved to the front of the prompt and broke the KV cache on every request

Other improvements

DeepSeek Harness launcher now falls back to npx when the global npm install fails, with Windows command-shim support

Onboarding flow has clearer introductory copy, a macOS header aligned with the native traffic-light controls, and Cmd/Ctrl zoom shortcuts disabled during onboarding so the fixed window keeps its intended scale

MLX dependency update (#17886)

Fixed broken default packaging caused by macOS-specific assumptions affecting Linux/Windows builds

Fixed the Apps header overlapping the macOS traffic lights during sidebar open transitions by synchronizing the header padding animation with the sidebar width animation

Full Changelog: v0.32.15...v0.33.0-rc2

What's Changed

• New desktop onboarding flow on first launch
• Caches resolved model metadata between requests, cutting time-to-first-token by roughly half (TTFT dropped from ~995 ms to ~524 ms in benchmarks)
• Fixes a bug where chat and generate could wedge after a mid-stream parser error
• Qwen 3.8 system messages are now normalized so non-leading system messages are handled consistently
• MLX and llama.cpp dependency updates

New Contributors

• @gaugarg-nv made their first contribution in #17752

Full Changelog: v0.32.14...v0.32.15

What's Changed

Models that don't set a repeat_penalty now default to 1.0 (off) instead of 1.1, matching other engines and speeding up speculative decoding; set a per-model parameter if an older model repeats itself.

Faster prefill on NVFP4 MLX models with a global scale, about 7–8% on Qwen3.6 and Muse Glimmer.

Fixed blob verification being skipped when an OCI manifest's config and layer share a digest.

New Contributors

@vigneshakaviki made their first contribution in #15504

Full Changelog: v0.32.8...v0.32.10-rc1

ModelOpt checkpoints apply a float32 global scale to every projection output on top of the per-group quantization scales. Running the multiply and the cast back to the activation dtype as separate eager ops costs an extra kernel launch and a materialized intermediate per projection.

Compile the multiply and cast into one kernel. On an M5 Max (medians of order-swapped A/B runs against main; greedy outputs byte-identical):

qwen3.6:27b prefill 703 -> 769 t/s +7.9%
muse-glimmer:30b prefill 790 -> 843 t/s +6.7%

Speculative decode is unchanged within noise on both models. Only checkpoints with a global scale are affected; single-scale nvfp4, mxfp8, and affine checkpoints take the unchanged path.

NVIDIA Nemotron 3.5 Lightning

NVIDIA Nemotron 3.5 Lightning is an open 30B mixture-of-experts (MoE) model with 3B active parameters built for that execution layer of always-on agents. It is designed for harnesses like OpenClaw and Hermes Agent – all supported by the NVIDIA NemoClaw open source security and management stack for running always-on AI agents.

ollama run nemotron-3.5-lightning

What's Changed

• Added the Nemotron 3 architecture
• Handle boundary condition in Muse Glimmer function calling parser

Full Changelog: v0.32.8...v0.32.9