Mistral just dropped the weights for its trillion-parameter model. As of October 29, you can download, run, and fine-tune what currently stands as the best open-weight coding model outside China — on your own hardware, with no API bill and no access that can be “cut off.” The name is Le Chonk. It earns it.
What You Are Actually Down #
Mistral Large 4 is a 1.05-trillion-parameter multimodal Mixture-of-Experts model. The MoE architecture is worth understanding before you assume “1T parameters” means what it sounds like. Only about 49 billion parameters fire per token — the rest stay dormant. You get the knowledge capacity of a trillion-parameter model at the inference cost of a ~50B model. That is the core premise, and it is why the hardware requirements, while serious, are not as absurd as the headline number implies.
The model handles text and images, supports a 1-million-token context window, and ships in three checkpoint formats: FP8 (the full-precision weights), NVFP4 (optimized for NVIDIA Blackwell GPUs), and GGUF quantizations ranging from IQ2_XXS to Q8_0 for broader hardware compatibility. All three are live on HuggingFace as of last week.
The Hardware Truth #
Here is where you need to be clear-eyed. Despite only 49B parameters being active during inference, all 1.05T parameters must reside in memory simultaneously. MoE routing requires all experts to be loaded before the model can decide which ones to activate.
At FP8 precision, you are looking at roughly 1,050 GB of VRAM. The minimum viable setup is an 8x H200 node — that configuration holds about 1,128 GB, leaving minimal headroom. A more comfortable deployment is a single DGX B200 node (1,440 GB). These are not rigs you have sitting in a home lab.
If you want something more accessible, the GGUF IQ2_XXS quantization brings the footprint down to around 233 GB — small enough for a Mac Studio with enough unified memory or two DGX Sparks. Quality drops with aggressive quantization, but it is a real option for experimentation. Day-zero inference support covers vLLM, llama.cpp, SGLang, and Hugging Face Transformers.
vllm serve mistralai/Mistral-Large-4-1T-A52B-NVFP4 \
--tensor-parallel-size 8 \
--enable-expert-parallel \
--dtype bfloat16 \
--max-model-len 131072
Benchmark Reality Check #
Mistral calls Le Chonk the best open-weight coding model, and on most benchmarks for self-hosted models, that holds. It scores 49.8% on the Artificial Analysis Coding Agent Index, ahead of DeepSeek V4 Pro and Qwen 3.8 Max. On Terminal-Bench 4.0, it hit 41% — a 13-point jump from the API preview and the current leader among open-weight models.
The honest caveat: closed models still win at the top. On DeepSWE v1.1, Le Chonk scores 61.7%, while Kimi K3 hits 68% and DeepSeek V4.1 Flash sits at 74.2%. This is not a closed-model killer. It is the best thing you can actually own and run, which is a different and arguably more interesting value proposition.
Token efficiency is the other headline stat. Le Chonk uses 47% fewer tokens per task than its nearest open-weight competitor, GLM-5.3, while matching it on benchmark scores. At scale, that difference translates directly to cost and latency.
Licensing: What You Can Build #
The Mistral Ecosystem Open-Weight License is a modified MIT license. You can use it, fine-tune it, and ship commercial products built on it. The one restriction: if you plan to operate it as a hosted service and sell API access, you need a separate commercial agreement with Mistral. For internal deployments, air-gapped environments, or products that use the model as a component rather than selling inference, you are clear.
Fine-tuning is well-supported. Mistral published LoRA examples using TRL and PEFT, targeting standard projection layers. If you want custom behavior without touching the base weights, that path is documented and working.
Why This Release Matters More Than Benchmarks #
Pierre Stock, Mistral’s VP of Science, put it plainly: “As a government or business, you’re able to have access to frontier intelligence to defend yourselves. Access that can’t be cut off because the model is open-weighted.”
That framing is not marketing. For regulated industries — finance, healthcare, defense — running a frontier coding model without sending data through a third-party API is a compliance necessity, not a preference. For European enterprises, the fact that this model was trained entirely on EU infrastructure on Mistral’s own hardware carries real weight under GDPR and NIS2.
At roughly $4,200 per month via API for 1,000 users at typical usage, there is also a straightforward financial case for organizations that can absorb the upfront infrastructure cost. The break-even math is real for anyone running inference at scale.
Le Chonk is available now in the Mistral Large 4 collection on HuggingFace. Mistral’s open weight release announcement has the full benchmark table and licensing terms. For inference setup, the technical guide on HuggingFace covers vLLM, llama.cpp, and SGLang configurations in detail. TechCrunch’s coverage has the original quotes from Mistral leadership.