A trillion parameters, downloadable
On October 29 — three weeks after putting Mistral Large 4 into public preview — Mistral released the model’s open weights. The model, which Mistral itself nicknames “Le Chonk,” is a natively multimodal mixture-of-experts system: roughly a trillion total parameters with only about 49 billion active per token (52 billion counting embeddings), a 1-million-token context window, and support for 160+ languages. Mistral’s headline for it: “Built for Coders, Trusted by Defenders.”
The weights are on Hugging Face as mistralai/Mistral-Large-4-1T-A52B — 54 safetensors files, which tells you everything about the hardware question before you even read the model card (Digital Soul Hub notes the October 29 “initial release” commit). Mistral says to deploy with vLLM or SGLang. A full-precision checkpoint of 1.05 trillion parameters is on the order of two terabytes — that’s my back-of-the-envelope math at 2 bytes per parameter, not a spec sheet figure — so this is data-center-class serving, not a homelab weekend project. A quantized NVFP4 checkpoint is also listed for multi-GPU servers, which brings it at least into the realm of serious lab hardware.
The catch in the license
This is the part worth reading twice. The weights are released under the Mistral Ecosystem Open-Weight license, a modified MIT license that — in Mistral’s own words from the announcement — leaves the model “free to use, modify, fine-tune, and commercialize, with a single condition: businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially.”
Self-host it for your own company or research? Fine. Wrap it in a hosted API and sell access? Talk to Mistral first. The announcement doesn’t precisely define where SaaS wrapping ends and model-as-a-service begins, so if your product is anywhere near that line, read the license and call them. For homelab and internal enterprise use, this is a non-issue — and that’s the point of the design.
What Mistral claims — and what’s actually independent
Every benchmark number in this section is Mistral’s own; independent validation will take weeks, and treat them accordingly. With that caveat, the claims are specific: 19.7% on Cognition’s private FrontierCode 2.0 (leading among open models), 48% on Artificial Analysis’s Coding Agent Index (tied with GLM-5.3 while using 47% fewer tokens per task), and 41% on Terminal-Bench 4.0 — up 13 points from the preview while cutting output tokens by 30%. On the security side: 51% on the Artificial Analysis Cyber Index, and a 96.6% refusal rate on malicious cyber requests, the highest Mistral says it has measured among open models. There are also three reasoning modes — Low, Medium, High (the default) — so you can trade performance for cost per request.
One genuinely independent data point is already in: Artificial Analysis assessed the open-weight release on October 29 and reported an Intelligence Index v5 score of 26 at high reasoning, 161.3 tokens/second, and $2.06 of compute per index task — down from $5.54 for the preview, a 63% drop in three weeks (via Guruji Corp’s analysis of the Artificial Analysis numbers). Note that index-version scores aren’t comparable across versions, so don’t stack that 26 against older v4 numbers.
If you’d rather just call it, hosted pricing starts at $0.70 per million input tokens, $0.07 cached, and $2.10 output. The model is served through Mistral Studio, the Mistral API, OpenRouter, NVIDIA NIM containers, and Microsoft Foundry — the first cloud provider to carry it.
Why this week matters for open weights
Le Chonk didn’t land in a vacuum. October has turned into open-weights season: Moonshot’s Kimi K3.1 Pro — a 2.8-trillion-parameter-class model Moonshot itself scored at 65.7% on Terminal-Bench 4.0 — is promised to publish its full weights by November 4, which Tech Times notes would make it “the most capable AI system ever freely distributed.” Moonshot took the opposite approach to Mistral on cybersecurity: it publicly documented removing offensive cyberattack training data from mid-training and RL for the open release (its own ablation shows a 9.5-point capability drop on its internal cyber eval), explicitly stating it doesn’t want the open weights to extend the frontier of offensive cyber capability. Meanwhile Google went the other direction entirely: Gemini 4 Carbon — announced October 28, which I covered in Friday’s roundup — is cyber-first but closed, going to trusted defenders before anyone else.
There’s also a sovereignty story here. Mistral trained Large 4 from scratch on European infrastructure — per Slashdot’s write-up of Mistral’s launch post (via AI Weekly), on 4,000 Nvidia Grace Blackwell GPUs over two months in European data centers — and the announcement is explicit that organizations can inspect, customize, and deploy it without depending on Mistral’s API. For a European lab, “download our trillion-parameter flagship” is a statement as much as a release.
The bottom line
Le Chonk is the most capable open-weight coding model you can download today, with an honest license, real (vendor-reported) benchmark gains, and one real independent confirmation that it’s gotten dramatically cheaper to run. You probably can’t self-host it on anything you’d call a homelab — but you can fine-tune it, inspect it, and serve it yourself at a scale where that matters. November 4 is the next date to watch: if Kimi K3.1 Pro’s weights land as promised, the open frontier gets two heavyweights in six days.