Mistral Large 4 ("Le Chonk") open weights: a 1T MoE sold on tokens per task, shipped in NVFP4
Sources: Mistral launch post · NVFP4 checkpoint · NVIDIA Dynamo recipe · @rohanpaul_ai on license and price · The Decoder · Hugging Face video. raw: raw/labs/2026-10-29-230224-labs.md, raw/rss/2026-10-29-the-decoder-paris-based-mistral-releases-le-chonk-a-strong-coding-m.md, raw/youtube/2026-10-30.json.
TL;DR. Mistral released the weights of an improved Mistral Large 4, a 1T-parameter, natively multimodal MoE with about 52B active parameters, trained from scratch on 3,800 Grace Blackwell GPUs in Europe. Three more weeks of RL since the 10-07 preview lifted Terminal-Bench 4.0 from 28% to 41% (first among open models, 1 point above GLM-5.3) while cutting output tokens 30%. On the Artificial Analysis Coding Agent Index it ties GLM-5.3 at 48% while using 47% fewer tokens per task. It leads Cognition's private FrontierCode 2.0 at 19.7%, scores 51% on the AA Cyber Index and 93% of Cybench, and refuses 96.6% of malicious cyber prompts. API price roughly halved to $0.70 / $0.07 cached / $2.10 per million tokens. Three reasoning modes (low, medium, high default). The license changed: a modified-MIT "Mistral Ecosystem Open-Weight" license that requires model-as-a-service providers to get a separate commercial license, tighter than Large 3's Apache 2.0. Independent read from The Decoder: on the overall AA index it still trails China's best (26 vs 32 for DeepSeek V4.1 Flash and GLM-5.3).
flowchart LR
W["Large 4 weights<br/><small>1T MoE, ~52B active</small>"] --> Q["NVFP4 checkpoint<br/><small>4-bit, day zero</small>"]
Q --> S["Dynamo + vLLM<br/><small>GB300 or Vera Rubin</small>"]
S --> SP["Speculative decode<br/><small>draft then verify</small>"]
SP --> O["Agent tasks<br/><small>47% fewer tokens</small>"]
W -.->|license| L["MaaS providers<br/><small>need separate license</small>"]
classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
classDef err fill:#ffe3e3,stroke:#e03131,color:#1b1b1b,stroke-width:2px
class W input
class Q,S core
class SP loop
class O exit
class L err
How it relates to prior wiki pages
- Confirms the token-efficiency-as-headline pattern: GPT-6.1 Terra (10-29) and Agent Arena's cost-per-task decomposition today both price capability per task, not per token. See LLM routing.
- Extends the open-cyber thread: Gemini 4 Carbon (10-29) went to defenders first; Mistral argues open weights are the defender channel. The same day Epoch AI reported open-weight GLM-5.3 built Firefox exploit chains for about $1,000 each (responsible-AI summary), which is the counter-case.
- License drift. A shift from Apache 2.0 to MaaS-restricted terms is a commercial response to hosts reselling open weights. Track whether others follow.