JetBrains open-sources Mellum 2.1: a 12B coding agent, 2.5B active
Apache-2.0 weights, 131K context, LiveCodeBench 82 — plus OpenAI's new Ultrafast tier, Claude Dashboards, and a Gemini Deep Research shutoff on Oct 23.
Copy markdown
A self-hostable coding agent that fits on one GPU
JetBrains open-weighted Mellum 2.1 — a 12B-total / 2.5B-active MoE under Apache-2.0 with a 131K context, built for agentic repo work (running commands, calling tools), not just autocomplete. It scores 82.0 on LiveCodeBench v6 and 47 on SWE-bench Verified, serves today on vLLM/SGLang/Transformers/Docker Model Runner (GGUF for Ollama + llama.cpp "coming soon"), and at 2.5B active params it's cheap enough to run a private coding agent against your own repo.
OpenAI adds an "Ultrafast" speed knob
The Responses API now accepts service_tier: "ultrafast" for GPT-6.1 Sol, giving latency-sensitive agents and UIs a faster path — open to all API users, with global processing and US/EU data residency. It's billed at a premium (reportedly around 6x standard rates), so reserve it for the interactive hops where speed actually changes the experience.
Claude's doc tools go GA; Dashboards and Motion arrive
Anthropic moved Docs, Slides, and Design out of beta and onto every plan including Free, and shipped two betas: Dashboards (live-connected to BigQuery, Databricks, and Snowflake) and Motion (code-based animated explainers, Team/Enterprise). If you've been generating decks or docs through Claude, the paywall is gone; analysts get dashboards wired straight to the warehouse.
Gemini's Deep Research preview shuts off Oct 23
Google deprecated deep-research-pro-preview-12-2025; it goes dark on October 23. Migrate to deep-research-preview-04-2026 (lighter, built for streaming to a UI) or deep-research-max-preview-04-2026 (more thorough) — if a pinned agent still calls the old ID, you have about 11 days before those requests break.