Mistral released the open weights for its 1-trillion-parameter Large 4 model on October 29th. Nicknamed “Le Chonk” — with no apparent irony — the model spent three weeks in closed preview before Mistral shipped FP8 and NVFP4 checkpoints to Hugging Face. Those three weeks were not idle: Terminal-Bench 4.0 scores went from 28% to 41%, and output token count dropped by 30%. This is a meaningfully better model than what reviewers benchmarked in October.
What You’re Getting #
Le Chonk is a natively multimodal Mixture-of-Experts with 1.05 trillion total parameters and roughly 52 billion active per token (49B routed experts plus embeddings and output layers). The architecture uses 128 routed experts with 4 active per token and a shared expert that fires on every pass. Context window is 1 million tokens — four times Mistral Large 3’s 256K — and the model handles text and image input natively via a 1.6B-parameter Pixtral-style vision encoder.
The model was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs inside Mistral’s own EU datacenters, which matters if your team has data-residency requirements that rule out US-based providers.
Preview vs. Open Weights: What Actually Improved #
The three weeks between preview (October 6) and open weights (October 29) were spent on reinforcement learning. The delta is real: Terminal-Bench 4.0 jumped 13 percentage points, from 28% to 41%. Output tokens dropped by 30%, which is an efficiency win that compounds quickly — fewer tokens means faster responses and lower API bills at the same quality level.
On the Coding Agent Index, Mistral Large 4 scores 49.8%, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max. It hits 61.7% on DeepSWE v1.1. It sits just behind GLM 5.3 on Terminal-Bench 4.0. If you benchmarked the preview and walked away unimpressed, the open weights version warrants another look.
The License — Read This Part #
Mistral ships Le Chonk under the Mistral Ecosystem Open-Weight License, a modified MIT. For most developers it functions exactly like MIT: use it, modify it, fine-tune it, ship it in your product. No restrictions on internal enterprise use.
The one carve-out: if you’re running Model-as-a-Service — meaning your business gives third parties access to model inference or further training — you need a separate commercial agreement with Mistral. This is the “MaaS clause.” Building an app that uses the model internally doesn’t trigger it. Selling API access to the model directly does.
Compare that to Kimi K3.1 Pro’s weights, which shipped with a $20M commercial license requirement. Mistral’s approach is considerably more permissive — but it’s not Llama- or Qwen-style unrestricted open source. If your business model is reselling inference, talk to Mistral before you build on this.
The Hardware Math #
The FP8 checkpoint weighs 1,060 GB. The NVFP4 checkpoint comes in at 610 GB. Neither is a single-GPU scenario. The reference deployment splits the model across eight GPUs — eight H200s or B200s for FP8, four GB300s for NVFP4. AMD MI350X with MXFP4 also works.
For teams without a GPU cluster, the only single-machine option is a Mac with 512 GB of unified memory running a 2-bit quantized version. Practical for evaluation; not for production agent workloads.
Deployment requires dedicated container images — standard vLLM and SGLang images don’t include Mistral Large 4 configuration. Use vllm/vllm-openai:mistral-large-4 or lmsysorg/sglang:mistral-large-4. The SGLang deployment docs cover the full setup.
If You’re Not Spinning Up Eight GPUs Today #
Mistral cut API pricing significantly when the open weights landed. Input tokens dropped from $1.36 to $0.70 per million; output from $4.18 to $2.10 — a 48% cut on input and 50% on output. At those prices, the API is competitive with mid-tier closed models.
The model is also available on OpenRouter for provider redundancy, and Mistral Studio has a web UI for testing. The Hugging Face model card is at mistralai/Mistral-Large-4-1T-A52B.
The Take #
Le Chonk is Mistral’s best model, and for the first time that’s a sentence worth saying without caveats. It’s genuinely competitive at the top of open-weight coding leaderboards. The license is permissive for direct use but has teeth for inference resellers. The hardware bar is high — this is a team-with-a-GPU-cluster story, not a hobbyist story.
If you’re evaluating open-weight models for a coding agent or long-context reasoning task and care about EU data sovereignty, this is now the benchmark to beat. If you’re building an inference product, read the MaaS clause before you commit.