Mistral releases Large 4 for download with cheaper API pricing

Mistral has published the weights of Mistral Large 4, its trillion-parameter model, so anyone can download and run it. The license has one notable limit. Companies that want to offer the model as a hosted service need a separate commercial license from Mistral.

Large 4 is a mixture-of-experts model, meaning it holds far more parameters than it uses for any one token. The Hugging Face model card describes 128 routed experts, with 4 active per token plus one shared expert. The card gives about 49 billion active parameters per token, or 52 billion when embedding and output layers are counted. Mistral's technical paper describes 52 billion active parameters. The model accepts text and image input and produces text output. It supports a 1M-token context window and offers three reasoning settings, low, medium and high, with high as the default. Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in its own European datacenters.

Mistral released a public preview earlier this month and says three more weeks of reinforcement learning produced the gains. On the Artificial Analysis Coding Agent Index, the blog reports a score of 48%, effectively tied with GLM-5.3, while using 47% fewer tokens per task on average. On Terminal-Bench 4.0 it reports 41%, up 13 percentage points from the preview, with output tokens reduced by 30%. The headline figure of 19.7% on FrontierCode 2.0 comes from Cognition's private benchmark, so outsiders cannot rerun it. The technical paper sets that score against 19.3% for DeepSeek V4.1-Flash 1017, on a single named benchmark rather than a field-wide ranking.

Much of the efficiency story depends on serving. Mistral's paper describes a calibrated NVFP4 checkpoint of 606 GB, which it says matches the FP8 reference on evaluations and can run on four GB300 GPUs. The checkpoint also ships with a draft head for speculative decoding. In that method, a small component proposes a block of three tokens and the main model checks them, so output is unchanged while throughput rises. Mistral measured up to 1.9 times higher throughput than serving the model alone, and it notes that actual gains depend on hardware and workload. The paper's serving results are about speed per request and per GPU, not about how many tokens an agent generates to finish a task.

The blog calls Large 4 the best open-weight model for coding, but Hugging Face reports 26 on the Artificial Analysis Intelligence Index, the highest among Western open-weight models. It says leading Chinese open-weight models score higher on that aggregate. The blog's human-preference results come from Surge Code, where Large 4 ranked behind only Claude Opus 5.5 among the models evaluated. The model is also served on the Mistral API as mistral-large-4. The Hugging Face page says official demos run on a small free budget per visitor during the first week of the release.