Mistral Large 4 Goes Open-Weight: 1 Trillion Parameters and What It Takes to Run Locally

Mistral AI released the open weights of Mistral Large 4, its 1-trillion-parameter flagship model nicknamed "Le Chonk," on October 29, 2026, three weeks after a public preview. Anyone can now download it from Hugging Face and run it on their own hardware under a modified MIT licence, with one commercial restriction for companies that resell it as a hosted service.

Mistral calls it the best open-weight model for coding, and it was trained from scratch in Europe, which matters to organisations that want to know where their AI comes from. But "open" does not mean "easy": this is a mixture-of-experts model that still needs a serious GPU server or a very large-memory workstation. Here is what was released, how it compares, and what it actually takes to run it locally.

Key facts at a glance

• Size: about 1 trillion total parameters, with roughly 49 billion to 52 billion active per token depending on how they are counted, and a context window of up to 1 million tokens.
• Formats: official FP8 and NVFP4 checkpoints, plus community GGUF quantisations for llama.cpp.
• Licence: the Mistral Ecosystem Open-Weight licence, free to use, modify, fine-tune and commercialise, except that model-as-a-service businesses need a separate commercial licence.
• API price: $0.70 per million input tokens, $0.07 cached and $2.10 per million output tokens.
• Focus: coding agents and defensive cybersecurity, with a reported 96.6% refusal rate on malicious cyber requests.

What Mistral released and why it matters

Mistral first showed Large 4 in public preview earlier in October. The company says three more weeks of reinforcement learning, guided by customer feedback, produced the version now on Hugging Face, with better coding, security and reasoning efficiency.

Under the hood it is a mixture-of-experts (MoE) model: instead of using all its parameters for every word, a router picks a few specialist sub-networks per token. Hugging Face's technical write-up says the text decoder has 72 layers, and 71 of them use MoE blocks in which the router selects 4 of 128 experts, plus one shared expert that always runs. That is why a 1-trillion-parameter model only activates around 50 billion parameters at a time. It also includes a 1.6-billion-parameter vision encoder, so it accepts images as well as text.

The model has three reasoning modes, low, medium and high, with high as the default. Reasoning cannot be switched off entirely, according to Hugging Face's documentation, so setting it to "none" still runs at high.

The strategic point is control. "Concretely, it means that, for instance, as a business company or as a government, you're able to have access to frontier intelligence to defend yourselves against cyberattacks. Access that can't be cut off by definition because the model is open-weighted," Pierre Stock, Mistral's vice president of science, said.

How it compares on coding and security

Mistral's headline claim is that Large 4 is the best open-weight model for coding. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with the leading Chinese open model GLM-5.3 while using 47% fewer tokens per task, which matters because tokens are what you pay for in compute or API fees. On Terminal-Bench 4.0 it improved by 13 percentage points over the preview to 41%, and on Cognition's FrontierCode 2.0 benchmark it leads with 19.7%.

Security is the other pillar. Mistral reports a 51% score on the Artificial Analysis Cyber Index, which measures whether a model can find, reproduce, validate and patch vulnerabilities without breaking the code. Within that index, Hugging Face highlights an 82% result on the specific reproduce-and-patch test, and 93% on Cybench's 40 security-competition challenges. On the B3 agent security benchmark, which tests resistance to attacks such as system-prompt exfiltration and malicious code injection, Mistral says the final model resisted 97.1% of attacks, up from 93.3% in preview.

All of these are vendor-reported or vendor-selected figures, so independent testing will matter.

What it takes to run it yourself

This is where expectations need resetting. Large 4 shipped with day-one support in Hugging Face Transformers, vLLM, SGLang and llama.cpp, so the software side is ready. The hardware side is demanding.

On a GPU server. Hugging Face's Inference Endpoints recipe uses 8 NVIDIA H200 GPUs running vLLM with the FP8 weights, serving up to 131,000 tokens of context and 16 requests at a time. It measured about 0.6 seconds to the first token on a 1,000-token prompt and around 65 tokens per second for a single request. The smaller NVFP4 checkpoint requires NVIDIA Blackwell-generation GPUs, and Dell's recipe runs it on four B300 GPUs.

On a workstation. The llama.cpp team publishes quantisations from 2-bit (IQ2XXS and Q2K) up to Q80. Quantisation shrinks a model by storing weights at lower precision, trading some accuracy for memory. Hugging Face says the 2-bit versions run on a single Apple Silicon Ultra machine with 512 GB of unified memory, where Q2K produced about 15 tokens per second on an M3 Ultra. On four RTX 6000 Pro cards, the NVFP4 build generated about 10 tokens per second at 32,000 tokens of context.

The basic llama.cpp command is a single line:

llama serve -hf ggml-org/Mistral-Large-4-1T-A52B-GGUF:Q2_K

Fine-tuning is also supported from day one through TRL, though Hugging Face's example uses a node of eight B300 GPUs. For most teams, a more realistic route is to use Large 4 to generate training traces for a smaller model, an approach Hugging Face suggests for security work such as patching a specific code base.

The caveats

Openness cuts both ways, and Mistral acknowledges it. During the preview the company monitored security-related traffic and said that while the model refused black-box attacks against live targets, "we see a potential for uplift to conduct real attacks," and that it is adjusting guardrails. A more specialised variant for penetration testing and red-teaming is limited to vetted partners such as Orange Cyberdefense.

The licence also deserves a careful read. It is described as a modified MIT licence, but businesses operating a model-as-a-service offering must obtain a separate commercial licence before using it commercially. Internal use, fine-tuning and building products on top are allowed.

Finally, the benchmark picture is partly self-reported. Mistral says a full technical report is "coming soon," and the community is only starting to run independent evaluations.

What it means for you

For developers and hobbyists, Large 4 is realistic only at 2-bit quantisation on an Apple Silicon Ultra machine with 512 GB of memory or on multi-GPU rigs. If you have neither, the API at $0.70 per million input tokens or a hosted endpoint is cheaper than buying hardware.

For companies in the UK, EU, Canada and Australia with data-residency rules, a self-hostable frontier-class model trained in Europe is a meaningful option. Mistral says it is also available through Microsoft Foundry, NVIDIA endpoints and OpenRouter if you want managed hosting.

For security teams, the pitch is a model that will help reproduce and patch vulnerabilities rather than refuse, deployed where your incident data never leaves your network.

What happens next

Mistral says Large 4 will be the base for a new family of specialised models and that a technical report is coming. Expect more community quantisations, independent benchmarks and fine-tunes on Hugging Face in the coming weeks, along with further guardrail updates.

Takeaways

• Check your hardware first: plan for 8 H200s, 4 Blackwell GPUs, or a 512 GB unified-memory Mac for 2-bit runs.
• Start with the API: test quality at $0.70 per million input tokens before investing in self-hosting.
• Read the licence: reselling the model as a hosted service needs a separate commercial licence.
• Treat benchmarks as claims: run your own evaluation on your code and tasks.
• Consider distillation: use Large 4 to generate traces that train a smaller, cheaper model.