Google DeepMind's Apache-2.0 EmbeddingGemma 2 (740M parameters, or 270M for text only) has an 8K context, 768-dim vectors you can truncate to 128, and jumps from 68.76 to 78.68 on MTEB Code, so it's a fit for local codebase indexing and coding-agent retrieval. Weights are on Hugging Face and Kaggle and run in sentence-transformers 6.1+, Ollama, llama.cpp, vLLM, MLX, and LM Studio.
The expanded Claude for Startups program, announced Oct 6 at SF Tech Week, includes up to five premium Claude Team seats for a year, $1,000 in API credits, Claude Marketplace access for building plugins, and office hours with Anthropic's Applied AI team. You qualify if your company was founded in the last five years or raised funding in the last two, and you apply through the Claude for Startups page…
Announced Oct 6 at Transcend, goal driven flows (GA this month) take an objective from the Duo CLI, headless mode, or Agentic Chat and carry it through review, tests, security scans, and approvals, while GitLab hosted GLM 5.3, Kimi K3, and MiniMax 3 are GA now with up to 8x more model calls per credit than frontier pairings. The MCP Server for outside coding agents also goes GA this month, and new credit caps can auto pause AI spend at a limit you set…
The Oct 5 release adds vllm preload , a daemon that keeps post quantized weights resident in GPU memory so engine restarts skip reloading, plus a big DeepSeek V4.1 Flash speed pass on Blackwell (SM100). Check your configs before upgrading: per request mm processor kwargs and media io kwargs are now rejected unless you pass trust request mm kwargs , tokenizer mode="slow" is gone, and online quantization="fp8" becomes fp8 per tensor…
Mistral opened a public API preview of Large 4, a multimodal MoE with 1.05T total / 49B active parameters and 1M context, and says open weights land Oct 27. Mistral admits it still trails frontier models on coding, and RL training isn't finished yet…
Turn it on under Settings → Permissions → Auto review (or approvals reviewer = "auto review" with approval policy = "on request" in config.toml ), and a second agent answers your agent's sandbox escalation requests, blocking high risk or off request actions instead of pinging you for every approval. It only reviews actions that already needed approval and doesn't widen the sandbox, but it's still a model that can be fooled, so keep your real limits in place…
Spotted something we missed? Start a thread.