Fireworks AI Release Notes

New model on Serverless: Rune 26B-A4B

Rune 26B-A4B is now available on Serverless. It’s a System One decision model, so it doesn’t reason and doesn’t generate text. You send a document and up to 256 typed questions about it, and it returns a calibrated probability for each answer from a single read. Questions can be multiple choice, true/false, or a graded scale. Output tokens are always zero, so you’re billed for input only.

Call accounts/fireworks/models/rune through the System One decisions API or the OpenAI-compatible Decisions API. It accepts text only, up to 262K context. Rates are on Serverless pricing.

Training SDK 1.2.30: clearer TITO errors and trajectory discard

fireworks-ai 1.2.30 and the updated Fireworks Cookbook complete the changes to how TITO sidecars handle failed agent turns, which began in 1.2.27. If you run agentic RL with TITO, upgrade both; the latest cookbook requires 1.2.30.

pip install --upgrade "fireworks-ai[training]>=1.2.30"

Errors identify their source

Before 1.2.27, the sidecar returned HTTP 502 for two failures that weren’t server-side: tito_model_malformed_output (unparseable model output) and tito_upstream_error (any sampling exception, including client timeouts and dropped connections). They are replaced by:

• tito_sdk_parse_error (422): the output couldn’t be parsed into an assistant or tool message. Includes a stable parser_reason_code, finish reason, and token counts.
• tito_server_http_error: the server returned an error, passed through with its real status, request ID, and retry details.
• tito_server_stream_error: the inference stream reported a failure, even after HTTP 200, so the turn is rejected before it can become training data.
• tito_transport_error (502, or 504 on timeout): the connection failed or timed out.

Responses also carry x-fireworks-error-origin and x-fireworks-error-code headers. Requests to a closing or closed trajectory return tito_trajectory_closed (410) with the reason it ended.

Unparseable trajectories are discarded

A parse failure makes the whole trajectory untrainable, including earlier successful turns. The cookbook RL loop drops it without retrying or stopping the run, and other rollouts continue. Parse failures don’t count toward the infrastructure circuit breaker, which now classifies transient failures more accurately and reports failure counts by reason. Watch producer/discard_errors_total: if most trajectories are discarded, a step can be left with no usable samples.

Failure traces are saved locally

Each failed trajectory leaves a private JSONL trace (request, prepared token IDs, raw completion, exception chain) in fireworks-tito-traces under the system temp directory, capped at 256 MiB. Traces from clean trajectories are deleted, and failures rejected by cookbook adapters use the same location and limits. To move or disable capture, pass failure_trace_config=TITOFailureTraceConfig(...) to TITOSidecar. If a trace can’t be written, the original rollout error is still reported. A trace shows whether the model, tokenizer, or renderer caused a parse failure; it contains raw prompts and completions, so review it before sharing.

Image inputs for FiretitanSamplingClient.sample

Since 1.2.27, sample and sample_async take a keyword-only images= list of base64 data URLs for token prompts with one placeholder per image. Prompts with unprojected image chunks are rejected so sampler-owned delimiters aren’t duplicated.

See the cookbook’s TITO failure debugging reference for every error code and how to read a trace.

Upcoming Serverless deprecation: GPT OSS 120B, Inkling, Nemotron, and US endpoints for MiniMax M3 and GLM 5.3 Flash

Several Serverless models will be decommissioned on November 7, 2026 to better serve newer, higher-performance replacements. This applies only to serverless inference. Dedicated deployments and Reserved Throughput are unaffected.

Action required

If you use any of the models below on serverless, migrate to a recommended replacement before November 7, 2026. After that date, they will no longer be available via serverless endpoints.

Recommended migrations

• GPT OSS 120B — move to a dedicated deployment of GPT OSS 120B
• Inkling — migrate to Muse Spark Open (US endpoint, 50% US regional premium waived)
• NVIDIA Nemotron 3 Ultra — move to a dedicated deployment of Nemotron Super VL
• NVIDIA Nemotron 3.5 Lightning 30B A3B — move to a dedicated deployment of Nemotron Super VL
• MiniMax M3 (US endpoint only) — migrate to DeepSeek V4.1 Flash 1017 (US endpoint). The global MiniMax M3 endpoint remains available.
• GLM 5.3 Flash (US endpoint only) — migrate to DeepSeek V4.1 Flash 1017 (US endpoint). The global GLM 5.3 Flash endpoint remains available.

GLM 5.3 and DeepSeek V4.1 Flash are not affected and remain available on serverless.

New models on Serverless

• DeepSeek V4.1 Flash 1017: accounts/fireworks/models/deepseek-v4p1-flash-1017
• DeepSeek V4.1 Flash 1017 (US endpoint): accounts/fireworks/models/deepseek-v4p1-flash-1017-us
• Muse Spark Open (US endpoint, 50% US regional premium waived): accounts/fireworks/models/muse-spark-open
• MiniMax M3.1: coming to serverless after its open-source release, expected Friday, October 23

If you want to switch to a dedicated deployment, see the Serverless model list and the on-demand deployment quickstart.

Serverless pricing update: DeepSeek V4.1 Flash

On October 1, 2026 at 00:00 UTC, serverless pricing for DeepSeek V4.1 Flash changes (uncached input / cached input / output price per 1M tokens):

• Standard: $0.22 / $0.007 / $0.66 → $0.30 / $0.006 / $1.20
• Priority: $0.275 / $0.00875 / $0.825 → $0.375 / $0.0075 / $1.50

This adjustment brings our pricing in line with current market rates for this model. It applies only to serverless usage. If you run DeepSeek V4.1 Flash on a dedicated deployment or use Reserved Throughput, your pricing is unaffected.

We are also rolling out infrastructure improvements designed to improve cache hit rate, minimize cost per task, and deliver a faster, more reliable experience across the board.

See Serverless pricing for the full rate card.

Managed RFT is paused; try RL on the Training API

Managed reinforcement fine-tuning (RFT) is paused. Managed Training no longer accepts new RFT jobs from the Fireworks UI, firectl, or the REST API. Existing jobs stay visible in your dashboard, and models you already trained with managed RFT keep serving.

Try RL on the Training API, where you write the rollout and training loop yourself and Fireworks runs the GPUs. Compared with managed RFT, you also get:

• Full-parameter RL on most current models, not just LoRA
• The training shape’s full context length, up to 524K tokens, instead of managed RFT’s fixed 32K limit
• Other methods such as on-policy distillation (OPD) and custom objectives

Your evaluator logic carries over. Start with Cookbook: Reinforcement Learning.

Managed SFT and DPO are unaffected.

Models no longer available for fine-tuning

The 70 models below were tunable only through managed RFT. With managed RFT paused, none of them can be fine-tuned on Fireworks anymore, and they no longer appear on the Models page. Inference on these models is not affected by this change.

Full list (70 models)

Serverless deprecation: DeepSeek V4 Pro (0813), DeepSeek V4 Flash (0731), and related models

The serverless deprecation announced for September 25, 2026 is now in effect. The models below are no longer available on public serverless, including Fast and US-only serverless endpoints where those existed. Dedicated deployments are unaffected.

Recommended migrations:

• DeepSeek V4 Pro (0813) — migrate to DeepSeek V4.1 Flash
• DeepSeek V4 Flash (0731) — migrate to DeepSeek V4.1 Flash
• DeepSeek V4 Flash Vision Exp — migrate to DeepSeek V4.1 Flash
• Muse Glimmer 30B — migrate to NVIDIA Nemotron 3.5 Lightning 30B A3B
• Kimi K2.6 — migrate to GLM 5.3 or Kimi K3
• Kimi K2.7 Code — migrate to GLM 5.3 or Kimi K3
• GLM 5.2, including GLM 5.2 Fast, GLM 5.2 Fast US, and GLM 5.2 US — migrate to GLM 5.3

See Serverless pricing and Which model should I use?.