Released Oct 28 by HKU NLP's DiffuSpace, the Qwen3.6-35B-A3B-derived MoE writes 32-token blocks in parallel instead of one token at a time, with a 128K context, a thinking mode, and SGLang and llama.cpp support. Its self-reported 73.6 on SWE-bench Verified, 51.7 on Terminal-Bench 2.1 and 680 tokens/s on one GB300 haven't been independently reproduced yet (Hugging Face).
Released Oct 27, the 1.73B text model has a 256K context, fits in about 1 GB at 4 bit, is post trained for local agent harnesses like Hermes Agent and OpenClaw, and fine tunes on a single GPU, with llama.cpp, MLX and vLLM support. Weights are on Hugging Face under the LFM Open License, free for commercial use only below $10M annual revenue, and Liquid says it isn't meant for agentic coding…
Announced Oct 28, commands that Copilot starts in JetBrains IDEs can now run in a GA local sandbox that limits filesystem, network and credential access by your or your org's policy. The same update brings Fusion, which replaces Auto and picks models per task at Standard, High or Max effort (30% off for 30 days), plus per model token usage when you hover an agent turn…
Released Oct 28 at Android Dev Summit, any Agent Client Protocol agent, including Claude Agent, Codex, Junie and Antigravity, can now run inside Android Studio with access to the emulator, build system, Logcat and Android skills, signing in with your own subscription or API key (Settings Tools AI Agents). The release also adds an AI Change Review for walking through large agent written diffs…
Announced Oct 28, the Agent Host now runs Copilot sessions outside any VS Code window, locally or on an SSH, dev tunnel or WSL machine, and keeps working after you disconnect as long as that host stays up. The open JSON RPC protocol lets you run code agent host and build your own client in TypeScript, Rust, Go, Kotlin or Swift (changelog)…
Announced at Universe on Oct 28, one cloud agent session now follows you across github.com, GitHub Mobile, Slack, Teams, the Copilot app, VS Code and Copilot CLI ( /sessions , Remote tab), running in a snapshotting sandbox with no one hour cap. On paid plans, sandbox compute now comes out of your included Actions entitlement, which becomes one dollar based pool shared with your builds…
Spotted something we missed? Start a thread.