On October 29, 2026, Mistral released the full open weights of Mistral Large 4, a model with about one trillion parameters that anyone can download from Hugging Face. Free to download is not the same as easy to run. The model card's own serving example uses eight GPUs, and the weights alone run to terabytes.
This is a weekend explainer on what the release actually gives you, what it costs through Mistral's API instead, and which of the headline benchmark numbers are vendor-reported. Mistral previewed the model on October 6; the October 29 release is an improved checkpoint.
Quick summary: Mistral Large 4 open weights were released October 29, 2026 on Hugging Face as mistralai/Mistral-Large-4-1T-A52B. About 1 trillion total parameters, about 49B active per token (52B with embeddings), 1M-token context, modified MIT license. Mistral's API charges $0.70 input and $2.10 output per million tokens. Benchmarks are vendor-reported.
What you get
Mistral describes a natively multimodal, 1-trillion-parameter model released under the Mistral Ecosystem Open-Weight license, which it calls a modified MIT license. Use, modification, fine-tuning and commercial use are allowed, but businesses running a Model-as-a-Service offering need a separate commercial license. (Source: Mistral AI)
Developers Digest adds the architecture: a mixture-of-experts design with 128 routed experts, 4 active per token, about 49B active parameters (52B counting embedding and output layers), a 1M-token context window, text and image input and text output. Online Tech Tips says a vision component of roughly 1.6 billion parameters handles images and documents, and that some outlets quote 1.05 trillion total parameters. (Sources: Developers Digest, Online Tech Tips)
The hardware reality
Developers Digest estimates about 2 TB of weights at 16-bit precision and says even 4-bit builds need a multi-GPU node; the model card's vLLM example uses tensor parallelism across 8 GPUs. Online Tech Tips reports Hugging Face's recipe runs the FP8 version on eight NVIDIA H200 GPUs, and an NVFP4 version on a four-GPU NVIDIA B300 server. The Developers Digest author did not run the command, since the model does not fit their hardware.
In plain terms: this is for companies with data-center GPUs, not for a gaming laptop. For everyone else, the practical route is Mistral's API or a hosted provider.
What it costs through the API
Mistral lists $0.70 per million input tokens, $0.07 per million cached input tokens and $2.10 per million output tokens. (Source: Mistral AI) A job with 10 million input tokens and 2 million output tokens would therefore cost $7.00 plus $4.20, or $11.20, at list prices before caching. That is our arithmetic from the listed rates.
Benchmarks: vendor-reported
These numbers come from Mistral's announcement and model card and have not been independently replicated, according to Online Tech Tips.
| Measure | Result | Context |
|---|---|---|
| Terminal-Bench 4.0 | 41% | Up 13 points from the October 6 preview (28%); vendor-reported |
| Cognition FrontierCode 2.0 | 19.7% | Highest among open models Mistral compared |
| Artificial Analysis Coding Agent Index | 48% | Effectively tied with GLM-5.3, with 47% fewer tokens per task |
| BrowseComp | 84.5% | Vendor-reported |
| Artificial Analysis Intelligence Index v5 | 26 | Highest among western open-weight models per the model card, which also says leading Chinese open-weight models score higher |
The last row matters. Developers Digest points out that the model card itself concedes the model trails the top Chinese open-weight systems on the general index, so the case for Large 4 rests on coding-agent strength, European provenance and licensing, not on topping every leaderboard.
Who should care
- Enterprises needing data control: running weights on your own servers keeps prompts in-house.
- Security and coding teams: Mistral targets agentic coding and cyber defense work.
- Hosting providers: the Model-as-a-Service clause means you need a separate license from Mistral.
- Everyone else: try it via API or a multi-model platform before investing in hardware.
Before buying GPU time, run the model on 10 of your own real tasks through the API and compare it with your current model. Vendor benchmarks rarely match your workload.
01When were the Mistral Large 4 weights released?
October 29, 2026, on Hugging Face as mistralai/Mistral-Large-4-1T-A52B. A preview came out October 6.
02Can I use it commercially?
Yes under the Mistral Ecosystem Open-Weight license, a modified MIT license. Companies selling it as a hosted service need a separate license from Mistral.
03What hardware do I need?
Data-center GPUs. The reported recipes use eight H200s for FP8 or a four-GPU B300 server for NVFP4.
04How much does the API cost?
$0.70 per million input tokens, $0.07 cached input and $2.10 per million output tokens, per Mistral.
05Is it the best open model?
Mistral claims the lead among open models on FrontierCode 2.0, but its own model card says it trails leading Chinese open models on the general Intelligence Index.
If you want to judge these tools on your own work, LumiChats puts 40+ premium AI models in one subscription for direct side-by-side comparison.
