Mistral Large 4 is Mistral's 1-trillion-parameter open-weight model, released as downloadable weights on October 29, 2026. It is built for coding agents and defensive cybersecurity, and it runs on your own hardware or through Mistral's API at $0.70 per million input tokens. This page checks whether your GPUs can run it and prices API use.
Last verified: November 1, 2026. Primary sources: Mistral release coverage via Reuters, Tech Startups, AlphaSignal, and Artificial Analysis scores. Independent explainer, not affiliated with Mistral.
Pick a checkpoint, your card, and how many you have. The estimate covers weights plus serving overhead only. Context length, KV cache, batch size, and concurrency need extra memory on top, so treat a close fit as a warning, not a pass.
Why 52B, not 49B? Both figures describe the same model. 49B is the routed active count per token; 52B adds embeddings and output layers. Pages that quote one without the other are not contradicting each other.
Memory estimates used here: NVFP4 about 560 GB, FP8 about 1,120 GB, weights plus overhead. Mistral says NVFP4 runs on as few as four B200 or B300 cards, which matches the NVFP4 estimate. FP8 needs roughly double, so plan on eight B200 class cards.
Mistral Large 4 on the API costs $0.70 per million input tokens, $0.07 cached, and $2.10 per million output tokens. Enter monthly volume and compare it with two other models at the rates used on this page.
Rival cached rates used here: Gemini 4 Carbon intro $0.10, GPT-6 Astra $2.50 per million tokens. Mistral's EU endpoint on OpenRouter bills 10 percent more ($0.77 / $2.31); the calculator uses Mistral's standard rate.
Mistral Large 4, nicknamed Le Chonk, is a natively multimodal mixture-of-experts model. It reads text, images, and code, works in more than 160 languages, and takes up to 1 million tokens of context in one pass. Mistral trained it from scratch on about 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, and positions it as Europe's answer to US closed models and Chinese open-weight models.
The practical pitch is control. Because the weights are downloadable, a bank, agency, or security team can run the model on its own infrastructure, fine-tune it, and keep sensitive data off a third party's servers. The tradeoff is the same one Mistral's own science lead states plainly: once weights are public, they cannot be taken back, and the same download is available to attackers.
| Spec | Value | Note |
|---|---|---|
| Total parameters | 1 trillion | Mixture of experts |
| Active per token | 49B (52B with embeddings) | 128 routed experts, 4 active, plus 1 shared expert |
| Context window | 1 million tokens | Serving memory and runtime limits apply |
| Languages | 160+ | Every official EU language included |
| Reasoning modes | Low, medium, high | High is the default; output token counts change with the mode |
| Serving | vLLM, SGLang, llama.cpp, Transformers | NVFP4 targets Blackwell GPUs; DSpark speculative decoding claims up to 1.9x throughput |
| Weights | Hugging Face: mistralai/Mistral-Large-4-1T-A52B | FP8 and NVFP4 checkpoints, plus day-one GGUF builds |
Large 4 first appeared on October 6 as an API-only preview. The October 29 release is a new event, not a re-announcement: the weights landed, the price moved, and the reasoning controls changed. Three more weeks of reinforcement learning sit between the two builds.
| Item | October 6 preview | October 29 release |
|---|---|---|
| Weights | API only | Downloadable, FP8 and NVFP4 |
| API price, input / output | $1.36 / $4.18 list (after a $0.68 / $2.09 promotion) | $0.70 / $2.10, cached $0.07 |
| Reasoning modes | Preview setup | Low, medium, high (high by default) |
| Terminal-Bench 4.0 | 28.3% | 41% |
| License | API terms | Modified MIT, with a hosted-service condition |
Price history matters for budgeting: this is the third rate the model has carried in 24 days. Do not lock a forecast to the preview promotion price.
Most headline scores come from Mistral. Artificial Analysis scores are independent measurements of the model, and they are labeled as such. No independent run of the final downloadable weights had been published at the time of writing, so treat vendor scores as claims to verify, not settled facts.
| Benchmark | Score | Source and reading |
|---|---|---|
| Terminal-Bench 4.0, agentic coding | 41% | Mistral-reported; up from 28.3% in the preview, with 30% fewer output tokens |
| FrontierCode 2.0, repository changes | 19.7% | Mistral-reported; top score among listed open-weight models |
| Artificial Analysis Coding Agent Index | 48% | Independent; tied with GLM-5.3, using 47% fewer tokens per task on average |
| Artificial Analysis Cyber Index | 51% | Independent; covers finding, reproducing, validating, and patching flaws |
| CyberGym-E2E-AA | 82% | Artificial Analysis; vulnerability reproduction and patching |
| Cybench, security competitions | 93% | Mistral-reported; solved challenges, not a general security grade |
| FinWorkBench | 73.3% | Reported; second among evaluated open models, behind Kimi K3 |
| Artificial Analysis Intelligence Index v5 | 26 | Independent; highest among Western open-weight models, behind leading Chinese open models |
Two cautions belong next to the cyber scores. First, Mistral reports a 96.6% average refusal rate on malicious cyber requests for the released configuration, but fine-tuning, prompts, and serving policy can change that behavior downstream. Second, a specialized penetration-testing variant stays limited to vetted enterprise design partners; the public checkpoint is not that variant.
The Mistral Ecosystem Open-Weight License is a modified MIT license. You can use, modify, fine-tune, and commercialize the model, including deployment on private infrastructure. The one condition that changes plans is for hosted services: a business that offers the model through a model-as-a-service product must get a separate commercial license from Mistral before doing so commercially. The announcement gives no revenue threshold and does not define the term in full, so read the license text in the Hugging Face repository before building a hosted product on it.
For everyone else, the license is the reason this release matters: it is the first time a 1-trillion-parameter Western model at this level can be run without asking a provider for continued access.
Yes, under the modified MIT license, but hardware and API use still cost money, and hosted-service businesses need a separate commercial license first.
Four B200 or B300 cards for the NVFP4 checkpoint, according to Mistral. The FP8 checkpoint needs roughly twice the memory, so plan on eight B200 class cards, plus extra memory for context and KV cache.
$0.70 per million input tokens, $0.07 cached, and $2.10 per million output tokens, from October 29, 2026. That is about half the list price that followed the preview promotion.
The weights became downloadable, the API price roughly halved, and low, medium, and high reasoning modes replaced the preview setup. Terminal-Bench 4.0 rose from 28.3% to 41%.
It is open-weight, not fully open source in the usual sense. You can download and fine-tune the weights, but the license adds a service restriction and no full technical report has been published yet.
Artificial Analysis scores it 26 on its Intelligence Index v5, the highest among Western open-weight models, behind several Chinese open-weight models. Most other headline scores are Mistral's own and await independent runs.
Benchmark figures are labeled Mistral-reported or Artificial Analysis in the table above. Where sources disagree on 49B versus 52B active parameters, this page uses both with the counting difference explained.
Machine-readable summary: mistral-large-4/llms.txt: specs, benchmarks, pricing, license, links. Updated 2026-11-01.
Ask an AI about this page:
ChatGPT · Claude · Perplexity · Grok · Gemini