Releasesby GlyphFireworks AI adds Rune 26B-A4B model to Serverless
Fireworks AI has added the Rune 26B-A4B model to its Serverless platform. This decision model processes documents and questions to return calibrated probabilities without generating text, and is billed based on input tokens only.
Also shipped Oct 27, 2026Fireworks AI in October 2026 →
OpenAI GPT-6.1 Terra is now available on Unity Gateway
Databricks now supports the OpenAI GPT-6.1 Terra model via Foundation Model APIs on Unity Gateway.
Editing files in Unity Catalog volumes is now generally available
Users can now edit files stored in Unity Catalog volumes directly within the Databricks workspace file editor, eliminating the need for manual downloads and uploads.
Released pg\_pinecone
The pg_pinecone extension is now available, enabling native vector search within PostgreSQL. It allows users to perform vector searches using the local engine or offload indexing to Pinecone for larger datasets.
[Launched] Generally Available: Cost management and tagging for Claude models in Microsoft Foundry
Microsoft Foundry now provides integrated cost management and tagging for Claude models within Azure Cost Management. This update ensures Claude usage is tracked under the specific Foundry resource rather than as general SaaS spend.
Gemini Robotics ER 2 (Preview)
Gemini Robotics ER 2 is now available in preview. This vision-language model is designed for robotics applications, supporting spatial reasoning, agentic code execution, and task orchestration.
xAI's Grok 4.7 is generally available
xAI's Grok 4.7 model is now generally available for production use on global and US multi-region endpoints.
More from Fireworks AI platformEvery Fireworks AI platform release →
Errors identify their source
Fireworks AI updated its error reporting to provide more granular source identification for failures, including new error codes for parsing and transport issues.
Unparseable trajectories are discarded
Fireworks AI now discards unparseable trajectories to prevent them from impacting training data quality.
Failure traces are saved locally
Failed model trajectories now generate private JSONL traces in the system temporary directory. These traces include request data, token IDs, and exception chains to help developers identify whether failures originated from the model, tokenizer, or renderer.
Image inputs for FiretitanSamplingClient.sample
The sample and sample_async methods now support a keyword-only images argument. This allows developers to pass base64 data URLs for token prompts using placeholders.
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.
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.
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.
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.