Playbooks · Coding tasks
Mistral Large 4 Le Chonk: a trillion parameters, open
Written November 5, 2026
Mistral's 1T-parameter model is open weight under a modified MIT licence with one condition. What it costs to run, and the Mac builds that fit.
What shipped
Mistral released the weights of Mistral Large 4, nicknamed Le Chonk, on October 29, 2026. It is a natively multimodal model with about 1 trillion parameters in total and 49 billion active per token (52 billion including the embedding and output layers), built from 128 routed experts with 4 active per token plus 1 shared expert. It has a 1-million-token context window, was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Europe, and covers 160+ languages. The model card is ungated: anyone can download the weights from Hugging Face.
The model has three reasoning modes — low, medium, and high, with high as the default — so a builder can trade reasoning depth for cost on simpler tasks. It also ships with a DSpark speculative-decoding draft head, which Mistral measures at up to 1.9x throughput. The weights come in two checkpoints: FP8 at 1.06 TB and NVFP4 at 610 GB.
The licence, in one condition
Le Chonk is released under the Mistral Ecosystem Open-Weight License, a modified MIT licence. The model is free to use, modify, fine-tune, and commercialize, with a single condition: a business that operates a model-as-a-service offering must obtain a separate commercial licence from Mistral before using the model commercially. There is no revenue threshold in the condition. If you run the model for your own work, or ship it inside a product that is not model-as-a-service, the condition does not apply to you.
What it costs to run
Through the Mistral API, the model costs $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens, under the model id mistral-large-4. It is also available on Mistral Studio, OpenRouter, and NVIDIA NIM. For a rough sense of what a workload costs at those rates, see the token cost calculator.
Self-hosting needs data-centre hardware. The vLLM recipe uses the dedicated image vllm/vllm-openai:mistral-large-4 with tensor parallelism across 8 GPUs. SGLang is the other supported serving route. Either way, the first cost is the download itself: 610 GB for the NVFP4 checkpoint, 1.06 TB for FP8.
The Mac builds
The community has built MLX versions of the model for Apple silicon. Three quantized builds exist: 2-bit at 338 GB, 3-bit at 459 GB, and 4-bit at 590 GB. The 3-bit build fits on a single 512 GB Mac. The 4-bit build needs two Macs, at about 300 GB of peak memory each. A full-precision BF16 copy would be 2.1 TB, so the quantized builds are the only practical route on a Mac.
The Mac builds run through mlx-lm pull request #2018, which loads the native FP8 weights as MXFP8. The community builds were tested with the development build mlx-lm 0.32.3.dev13, not the released 0.32.2, so expect to install from the development build or the pull request branch. The pull request covers the text model only: no vision support and no DSpark draft head yet.
Benchmarks
Every figure below is Mistral's own reporting, not an independent measurement. Mistral calls Le Chonk the best open-weight model for coding, and reports:
- FrontierCode 2.0: 19.7%, leading the open-weight models on Cognition's private benchmark.
- Artificial Analysis Coding Agent Index: 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task on average.
- Terminal-Bench 4.0: 41%, up 13 points from the public preview.
- Artificial Analysis Cyber Index: 51% on defensive security work.
- Intelligence Index v5: 26, per the model card.
- Human preference: in Surge Code's blind pairwise evaluations, Le Chonk was preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models evaluated.
Take the coding figures as a signal that the model is competitive with the best open-weight options, not as proof it beats a closed model on your workload. The token-efficiency claim matters more than the headline scores for agent work: fewer tokens per task is what keeps a long coding session cheap.
Limits
- Running the weights locally starts with a 610 GB to 1.06 TB download, and serving needs 8 data-centre GPUs. The API route is the only cheap way to try the model first.
- The MLX builds are community work, not Mistral releases. The pull request covers the text model only, and the builds need a development version of mlx-lm.
- The benchmark figures are Mistral's own numbers. The technical report is linked below for anyone who wants the detail behind them.
- The licence condition binds only model-as-a-service businesses. It does not restrict running the model, fine-tuning it, or shipping it inside another product.
Related playbooks
- Judge Muse Spark Open before you download it: 816B weights, a custom license, and a cluster-sized hardware bar
- Mistral Studio: managed agents, in preview
- Mistral Vibe 3.1.0: see what fills the context
- Open an untrusted repo without running code
Sources
- Mistral Large 4 (Le Chonk) is now open weight — Mistral AI: weights release, benchmarks, licence condition, API price
- Mistral-Large-4-1T-A52B model card — Hugging Face: architecture, context window, training hardware, community MLX builds
- Add Mistral Large 4 text model — mlx-lm pull request #2018: native FP8 load as MXFP8, 3-bit build on one 512 GB Mac
- Token cost calculator — Agent Upstart: estimate a workload at the API rates above
Back to: Coding playbooks hub · Also: Token cost calculator