Kimi K3.1 Pro open weights ship: 2.8T MoE, 104B active, 1M ctx
Moonshot's first K3.1 model self-hosts now with near-Opus agent scores — plus Liquid's on-device LFM3, Copilot in JetBrains Air, and a GitHub Actions lockdown.

Copy markdown
The weights are live on Hugging Face
Moonshot's first K3.1-series model is an open-weight, natively multimodal MoE — 2.8T total parameters, 104B active, 1M-token context. Pull it today with `vllm serve moonshotai/Kimi-K3.1-Pro` or SGLang, or hit the OpenAI/Anthropic-compatible hosted API at platform.kimi.ai (model id `kimi-k3.1-pro`).
Near-Opus on agents, not on raw reasoning
Moonshot's own numbers put it at Terminal-Bench 4.0 65.7 (Opus 5.5: 66.4) and BrowseComp 92.1, edging out Opus 5.5 and GPT-6 Astra. The catch: HLE-Full is 51.1 vs Opus's 59.8 and it trails on OSWorld — a coding and browsing workhorse, not a reasoning champ. All self-reported.
Liquid's LFM3-1.7B runs a real agent on your laptop
Liquid AI shipped a tri-hybrid 1.7B text model built for on-device agents: ~1GB at 4-bit, 256K context, 16 languages, and claimed 2.5x faster prefill / 1.8x faster decode than peer small models. Ships as GGUF, MLX 4-bit and ONNX, and scores well on tool-calling (BFCL V4 71.4) and instructions (IFEval 96.7) — just not agentic coding.
Copilot lands inside JetBrains Air
GitHub Copilot is now selectable as an agent in JetBrains' new Air IDE, running on your existing Copilot subscription. Cloud tasks execute in isolated VMs billed to your Copilot plan — not JetBrains AI credits — so there's no second meter to watch.
Heads-up: GitHub blocks pull_request_target Dec 2
If your public-repo Actions use `pull_request_target`, they'll start failing: brownouts Nov 9 and Nov 16 (16:00–18:00 UTC), then full enforcement Dec 2. Audit affected workflows in Insights now and add an explicit Actions event policy to keep the ones you need. Private and internal repos are unaffected.