Mistral Large 4 open weights went up on Hugging Face on Thursday, October 29, three weeks after the Paris lab launched the model as an API-only preview. The trillion-parameter model it nicknames Le Chonk is now a roughly 1,060 GB download you can fine-tune and run yourself.
The same day, Mistral set the API price at $0.70 per million input tokens and $2.10 per million output. That replaces the “sale” and “list” prices we reported on October 12, so the bill won’t double when a promotion ends.
- Price: the weights are free to download. Mistral’s API charges $0.70 per million input tokens, $0.07 cached and $2.10 output.
- Where: Hugging Face (FP8 about 1,060 GB, NVFP4 about 610 GB), Mistral Studio, OpenRouter, Microsoft Foundry and Nvidia’s hosted endpoints.
- Hardware: server class. Mistral’s own vLLM example spreads it across eight GPUs.
- The catch: model-as-a-service businesses need a separate commercial license.
What the Mistral Large 4 open weights include
The model card describes a mixture-of-experts model with about 1 trillion total parameters (Mistral’s docs say 1.05 trillion), 128 routed experts with 4 active per token plus one shared expert, and 49 billion active parameters per token, or 52 billion counting embeddings and output layers. It reads text and images, writes text, holds a 1 million-token context window and has three reasoning settings: low, medium and high, which is the default.
There are two official checkpoints. The FP8 repository’s weight files add up to about 1,060 GB on Hugging Face. A quantized NVFP4 version is about 610 GB and targets Nvidia’s Blackwell GPUs. Mistral doesn’t publish a minimum hardware spec, but its recommended vLLM command uses a tensor-parallel size of eight. Unsloth and the llama.cpp team have published smaller GGUF quantizations.
The new price, and what happened to the sale
When we checked Mistral’s docs on October 12, the preview showed a sale price and an “original” price twice as high. The model page now shows a single rate, which matches the “starting today” price in the release post. Per million tokens:
| Mistral Large 4 API (per 1M tokens) | Input | Cached input | Output |
|---|---|---|---|
| Preview sale price (Oct 12 check) | $0.68 | $0.07 | $2.09 |
| Preview list price (Oct 12 check) | $1.36 | $0.14 | $4.18 |
| From October 29 | $0.70 | $0.07 | $2.10 |
The real price landed a couple of cents above the old sale rate and about half the old list rate. OpenRouter lists the same $0.70 and $2.10. One oddity: the docs page still tags the model “Public Preview.”
The license has one real condition
The Mistral Ecosystem Open-Weight License is a modified MIT license that allows commercial use, with one exception. A company running a “Model as a Service” business needs a separate commercial license from Mistral. The license defines that as giving a third party inference or post-training access, for example through an API, with “meaningful control over the inputs, parameters, or training data.” Internal use is exempt, and so are end-user products where the model sits inside a specific feature. An inference host reselling Le Chonk by the token needs Mistral’s sign-off. A company running it behind its own app probably doesn’t.
Benchmarks: Mistral’s claims and one independent check
Mistral says three more weeks of reinforcement learning made Large 4 “the best open-weight model for coding.” By its own charts, the model scores 48% on Artificial Analysis’s Coding Agent Index, which Mistral calls “effectively tied” with China’s GLM-5.3 while using 47% fewer tokens per task. It reports 19.7% on Cognition’s private FrontierCode 2.0 test and 41% on Terminal-Bench 4.0, up from 28% for the preview. In Surge’s blind coding comparisons, graders preferred it over GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, but it ranked behind Claude Opus 5.5.
On security, Mistral claims 51% on Artificial Analysis’s Cyber Index and a 96.6% refusal rate on malicious cyber prompts. It also admits that in live traffic since the preview, it sees “a potential for uplift to conduct real attacks” and is “modifying our guardrails accordingly.” With the weights public, those API guardrails don’t follow the file.
The independent Artificial Analysis page puts Large 4 at 26 on its Intelligence Index v5, above the median of 23 for comparable open models, at 161 output tokens per second on Mistral’s API. That’s well behind the closed frontier: Gemini 4 Carbon scored 59 on the same index this week. Mistral’s own model card concedes that leading Chinese open-weight models score higher on the aggregate.
What to watch next
Mistral says a full technical report is “coming soon.” It should show how the cyber and refusal numbers were measured. The nearer date is November 4, when Moonshot’s Kimi K3.1 Pro weights are due. That will be the first fresh challenge to Le Chonk’s open-weight coding claims.
Frequently asked questions
Are Mistral Large 4’s weights open?
Yes. Since October 29, 2026, the weights have been on Hugging Face in FP8 (about 1,060 GB) and NVFP4 (about 610 GB) versions under the Mistral Ecosystem Open-Weight License. Commercial use is allowed, but companies running a model-as-a-service business need a separate license from Mistral.
How much does Mistral Large 4 cost?
Downloading the weights is free. On Mistral’s API it costs $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens, as of October 30, 2026. That single price replaced the preview’s sale and list prices.
What hardware do I need to run Mistral Large 4?
Mistral doesn’t publish a minimum spec. Its recommended vLLM setup uses eight GPUs, the FP8 weights are about 1,060 GB and the NVFP4 version, built for Nvidia Blackwell GPUs, is about 610 GB. Smaller community GGUF quantizations exist, with quality trade-offs Mistral hasn’t measured.
Is Mistral Large 4 better than GLM-5.3 or Kimi K3?
Mistral’s charts show it tied with GLM-5.3 on Artificial Analysis’s Coding Agent Index while using fewer tokens, and preferred over GLM-5.3 and Kimi K3 in Surge’s blind coding tests. On Artificial Analysis’s broader Intelligence Index, Mistral says leading Chinese open-weight models still score higher.
Sources
- Mistral, “Mistral Large 4 (Le Chonk) is now open weight” (October 29, 2026; price, benchmarks, guardrails, availability): mistral.ai
- Mistral docs, Mistral Large 4 model page (single $0.70/$0.07/$2.10 price, “Public Preview,” v26.10, 1M context): docs.mistral.ai
- Hugging Face, mistralai/Mistral-Large-4-1T-A52B model card and file list (architecture, vLLM settings, FP8 size): huggingface.co; NVFP4 checkpoint: huggingface.co
- Mistral Ecosystem Open-Weight License: huggingface.co
- Artificial Analysis, Mistral Large 4 (High) model page (Intelligence Index v5, speed, cost per task): artificialanalysis.ai
- OpenRouter, mistralai/mistral-large-4-0 listing (price): openrouter.ai
- Mistral AI on X, release post (October 29, 2026, 10:21 AM ET): x.com
- Our October 12 preview report (sale and list prices): roboticaguide.com/
All sources checked on the morning of October 30, 2026 (ET). File sizes are the sum of the weight files listed on Hugging Face. Benchmark figures are Mistral’s or Artificial Analysis’s, as labelled.
Last updated: October 30, 2026. To report an error, see our corrections page. Articles are drafted with AI assistance and reviewed and edited by an editor; see our editorial policy.
