Mistral published the weights for Le Chonk on October 29, three weeks after it showed the model to anyone (Mistral). Le Chonk is Mistral Large 4, a one-trillion-parameter mixture-of-experts model with 49 billion parameters active per token, and Mistral describes it as leading the open-weight benchmarks in coding and defensive security (Mistral).
The number that matters for anyone planning a deployment is not in the benchmark table. It is in the serving recipe. Mistral’s model card documents a vLLM command with a tensor-parallel size of eight, and the smallest published GGUF build for llama.cpp is 298 GB (model card, GGUF repository). The same card settles the question for most people in one line: “For optimal performance, we recommend using the Mistral AI API if local serving is subpar” (model card).
An open licence, a model well past consumer hardware, and a hosted API priced per token. That is what “open weight” means in 2026, and the gap between the licence and the hardware bill is where the decisions get made.
What actually shipped
The weights sit on Hugging Face under the Mistral Ecosystem Open-Weight License, a modified MIT licence. Mistral states the terms plainly: the model is “free to use, modify, fine-tune, and commercialize, with a single condition: businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially” (Mistral).
The model card describes roughly 1 trillion total parameters with about 52 billion active including embeddings and output layers, built from 128 routed experts with four active per token plus one shared expert. It takes text and image input through a Pixtral-style vision encoder, carries a 1M-token context window, and is natively fluent in 160+ languages (model card).
Simon Willison reported the October 6 preview with the note that “Mistral are back in the game”, and that it was “available via their API” with the open weights promised for the end of the month (Simon Willison). Mistral attributes the change to three weeks of additional reinforcement learning, which “propelled it to the forefront of open-weight coding” (Mistral).
The scoreboard moved accordingly. Mistral Large 4 reaches 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task on average, and it improved by 13 percentage points to 41% on Terminal-Bench 4.0 against the public preview, taking the lead among open-weight models while reducing output tokens by 30%. On Cognition’s private FrontierCode 2.0 benchmark it leads with a score of 19.7% (Mistral).
Security numbers follow the same pattern: 51% on the Artificial Analysis Cyber Index, a 96.6% refusal rate on malicious cyber requests, which Mistral calls “the highest we measure among open models”, and 97.1% attack resistance on the B3 Agent Security Benchmark against 93.3% in the preview (Mistral).
One counterweight matters. On the Artificial Analysis Intelligence Index v5, Mistral Large 4 scores 26, “the highest among western open-weight models”, and the card concedes that “the leading open-weight models from China score higher on this aggregate index” (model card). The card also records that the model was trained from scratch on 3,800 NVIDIA GPUs in Mistral’s own datacenters in Europe (model card).
The hardware a trillion parameters needs
The documented serving path is a node of eight GPUs. The vLLM recipe passes --tensor-parallel-size 8, the SGLang recipe passes --tp-size 8, and the card annotates that flag as an “example tensor-parallel size for a node of 8 GPUs; adjust to your topology”. The advice for a tighter box is to “lower --max-model-len and/or --max-num-seqs” (model card).
The llama.cpp route exists, and the quantisation ladder is published in full: BF16 at 2.10 TB, Q8_0 at 1.12 TB, Q4_K_M at 635 GB, Q2_K at 381 GB and IQ2_XXS at 298 GB, with the two smallest marked as fitting in 512 GB of unified memory (GGUF repository). The repository’s quickstart is llama serve -hf ggml-org/Mistral-Large-4-1T-A52B-GGUF:IQ2_XXS, and the model page lists Ollama, LM Studio and Jan among the applications that can open these builds (GGUF repository).
So the honest local hardware floor is a machine with 512 GB of unified memory for the most aggressively quantised builds, and nothing smaller. Every step up the ladder lands in datacenter territory, the same kind of wall Kimi K3 ran into.
Two additions make that path less painful. The checkpoint ships an embedded DSpark speculative decoding draft head, and Mistral measured “up to 1.9x higher throughput than serving the target model alone”, with gains that “depend on hardware and workload”. A separate NVFP4 checkpoint is available with a smaller memory footprint and the same performance, supporting vLLM and SGLang (model card).
What it costs to use it
Mistral prices the hosted route at $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens (Mistral). Token efficiency matters most in practice: Mistral says the model uses 47% fewer tokens per task than GLM-5.3 on the Coding Agent Index, and TechCrunch reports the same claim while framing ML4 as comparable to GLM-5.3 on cyber work (Mistral, TechCrunch).
That is the case for open weights as Mistral tells it. Pierre Stock, Vice President of Science at Mistral, told TechCrunch that open weights mean “Access that can’t be cut off by definition because the model is open-weighted”, and that organisations can “deploy this model on-prem in their own ecosystem” (TechCrunch). Mistral’s CEO Arthur Mensch frames the same idea as AI that should “be a tool for empowerment, not dominance” (TechCrunch).
The licence condition shows where the control sits. A company with the hardware cannot spin up a hosted service on these weights without a separate commercial licence from Mistral. The guarantee being bought is independence from a provider cutting access, not independence from a price list.
What this means
Open weights and local execution are two different promises, and this release separates them cleanly. The licence is genuinely open, the weights are downloadable, and a competent team with eight GPUs can run them. The smallest build that fits on a single machine needs 512 GB of unified memory, and Mistral’s own card recommends the API when local serving is subpar.
That is not a flaw in the release. It is the shape of the market. The models with this capability are trained and served at datacenter scale, and the open-weight path suits organisations with hardware, compliance requirements or a reason not to send code to an API. For everyone else, the useful question is not whether the weights exist. It is what the tokens cost when you rent the capability instead.
The bottom line
Le Chonk is a real open-weight release with a permissive licence, a strong coding and security benchmark record, and an honest limitation written into its own documentation: the official recipe assumes eight GPUs and the smallest quantisation is 298 GB. Judge it as what it is, a frontier model for people who have the hardware, and price the API at $0.70 per million input tokens and $2.10 per million output tokens as the practical route for everyone else.