Mistral ships Le Chonk's open weights: 1T params, runs on vLLM now
The improved checkpoint trails only Opus 5.5 in blind coding tests. Plus: Grok's API now watermarks text by default, and Buoyant open-sources Kinference.

Copy markdown
A trillion-param open model, free to self-host
Mistral's full Le Chonk weights (Mistral-Large-4-1T-A52B — 1T total, ~52B active, 1M context) are public under a modified-MIT license that permits commercial use, with vLLM, SGLang and NVIDIA NIM recipes ready to serve. The improved checkpoint scores 48% on the Artificial Analysis Coding Agent Index at 47% fewer tokens and trails only Claude Opus 5.5 in Surge's blind human coding test; the hosted API runs $0.70/$2.10 per 1M input/output.
Grok's API now watermarks your text by default
Since Nov 7, output from seven Grok models (grok-4.7 down to grok-4.3 plus the grok-4.20 family) carries an invisible statistical watermark — read the x-watermark response header to tell if a response is marked. Only team admins can opt out, per model and only after asserting a legal exception; xAI is racing the EU AI Act's Article 50(2) deadline in December.
Buoyant open-sources a control plane for vLLM
Kinference (Apache-2.0) routes and load-balances across vLLM endpoints using KV-cache events and engine metrics, speaking OpenAI chat-completions over gRPC and configured via Kubernetes CRDs. Buoyant claims up to 95% lower median time-to-first-token versus plain K8s serving — but it's alpha, with no rate limiting, /v1/messages or multi-replica support yet.