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.
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…
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…
Launched Oct 6 by an indie developer, the persistent world lets agents join over MCP ( claude mcp add transport http agentonia https://mcp.agentonia.com ) or REST to work, trade, start companies, run for office, and pass laws across four regions with different governments, with founder slots capped at 100. Spectators can read every agent's private messages and reasoning, so don't send in an agent that has access to anything you'd mind leaking…
The server at https://anaconda.com/api/mcp now runs remotely and only returns data, giving Claude Code, Cursor, Codex, Kilo, and other remote MCP clients package info and forum search on any Anaconda account, with package security and org policy tools on Business and Enterprise tiers. Unlike the beta it can't install or change anything on your machine, so your agent still runs conda itself under your own permission settings…
The RC of Java and Kotlin by IntelliJ IDEA brings refactorings, DAP debugging, and Maven/Gradle/Bazel support to VS Code based editors, and in Cursor the agent can use the same language server for semantic search, navigation, and analysis. It's free for 30 days under the EAP, but after the stable release (due soon) you'll need an IntelliJ IDEA Ultimate subscription once a 30 day trial ends…
Spotted something we missed? Start a thread.