Mistral Large 4: The 1T Model You Can't Run at Home

Mistral AI put out a preview of its biggest model yet on October 6: Mistral Large 4, nicknamed "le Chonk." It is a natively multimodal mixture-of-experts model with about 1 trillion parameters in total and 52 billion active per token. The weights are not out yet. Mistral says they land by the end of October.

The headline numbers come with a catch that matters for anyone running models at home. The 52 billion active parameters decide how fast the model runs, not how much memory it needs. All one trillion weights have to sit in memory, which puts Large 4 at roughly half a terabyte for the weights alone at Q4 quantization, before you generate a single token. For comparison, GLM-5.3-Flash is 320B total parameters, which still works out to about 180GB at Q4, so it lives in data centers too. Large 4 is a multi-GPU cluster job, not a desktop.

Mistral knows this. The company trained the model from scratch on 3,800 of its own Grace Blackwell GPUs in European data centers, and the preview runs on that same hardware. For now, Large 4 is an API you call, not a checkpoint you download. The open-weight release is the part that makes it interesting for self-hosters: you could run it under your own policies, fine-tune it, and audit it, once the weights and the license text exist.

The interesting part is the shape of the release

Large 4 is a hybrid instruct-and-reasoning model, which means one model can answer directly or think step by step on hard tasks, instead of two separate models. Mistral is aiming it at agentic coding: it scored 61.7 percent on DeepSWE v1.1, a benchmark for agentic software engineering, slightly ahead of GLM-5.3.

There is a second thing worth watching. Mistral says it is red-teaming the model with cybersecurity partners and state authorities before the weights ship, and that those partners get the model with reduced moderation and expanded cyber capabilities. Whether you like that arrangement or not, it is a sign of where the open-weight frontier is going: the biggest models are dual-use by default, and the release process now includes a security review.

What you can do with this today

Try the preview through Mistral's API. Run your own prompts against it, especially on the agentic coding tasks it claims to be good at, and compare cost and latency against what you run now. Vendor benchmarks are a starting point, not a verdict.

While you wait for the weights, run the models you can fit. A 24B dense model at Q4 needs about 14GB of VRAM. A 32B needs about 18GB. If you want to know exactly what your card can hold, run the numbers in the VRAM calculator before you download anything.

One thing to check before you build on it

Open-weight is not the same as open-source. The license terms were not public when this was written. Check the license text on release day before you put this model in a product, especially for commercial use and redistribution. That is the unglamorous step that decides whether Large 4 is a research toy or infrastructure.