Models
Mistral releases Large 4's weights: what developers need to know
Mistral Large 4's weights are available. We explain its license, API pricing, coding claims and demanding hardware requirements.
Mistral Large 4, nicknamed Le Chonk, is now available as downloadable weights. The release follows the model's public API preview on October 6 and gives developers another way to deploy or customize Mistral's largest model. The weights are open, but their license has conditions, and running a model of this size still requires substantial hardware (Mistral model card; Hugging Face).
Mistral positions Large 4 for coding, agents and cybersecurity as well as general text and image understanding. Here's what has changed, how much the hosted API costs, and what to consider before running it yourself.
What is Mistral Large 4?
Large 4 is a multimodal mixture-of-experts model: it accepts text and images and produces text. Mistral describes approximately one trillion total parameters and 49 billion active parameters per token, or 52 billion when embeddings and output layers are included. Its published context window is one million tokens (model card).
A mixture-of-experts architecture activates only part of the network for each token. That does not mean the entire model needs only enough memory for its active parameters: the full set of weights still has to be stored and made available to the serving system.
The release includes FP8 and NVFP4 checkpoints; community GGUF quantizations are also available. Mistral lists low, medium and high reasoning efforts, with high as the default. Performance depends on the checkpoint, serving setup, context length and workload (Mistral; Hugging Face).
Open weights, with a specific license
Mistral distributes the model under the Mistral Ecosystem Open-Weight License, a modified MIT-style license. Its model card says commercial and noncommercial use are permitted, while operating a model-as-a-service business requires a separate commercial license from Mistral (model card).
That makes the weights downloadable and usable under the stated terms; it does not make the license identical to Apache 2.0 or unrestricted MIT. Developers should read the actual license for their deployment, especially if offering inference to customers.
Coding and cybersecurity: what the scores say
Mistral reports a score of 48 on the Artificial Analysis Coding Agent Index and says the model used 47% fewer tokens per task than GLM-5.3 in that comparison. Those figures are reported by Mistral and depend on the benchmark's specific agent configuration and workload (model card).
Mistral also emphasizes security use cases, such as finding and correcting vulnerabilities. Hugging Face cites a reported 82% on one vulnerability reproduction-and-patching test. That is a score on one task within a broader evaluation, not proof that the model can secure a codebase on its own (Hugging Face).
Different published materials cite different coding-index figures and may reflect different evaluations or versions. Until the exact runs can be matched, we would not combine them into a single ranking. For a real project, compare models on the same representative tasks, record the configuration and review the resulting code.
What does the API cost?
Mistral's model page lists the following prices per million tokens as of October 29, 2026:
| Token type | Listed price |
|---|---|
| Input | $0.70 |
| Cached input | $0.07 |
| Output | $2.10 |
Source: Mistral documentation, checked October 29, 2026. The page still labels the service a public preview. Pricing and availability may change.
For illustration, one million uncached input tokens plus one million output tokens would cost $2.80 at those listed rates. Actual agent tasks can make multiple calls and consume very different amounts of input, output and cached tokens. These API prices are separate from the cost of running downloadable weights yourself.
Can you run it on a regular laptop?
Not realistically with the configurations documented so far. Hugging Face describes a highly quantized version running on an Apple Silicon Ultra machine with 512 GB of unified memory; other configurations use multiple GPUs. The exact requirement depends on quantization and runtime, but this is a very large model (Hugging Face).
For a developer with a typical Mac or PC, a hosted API is the practical starting point. Downloadable weights remain valuable to organizations with the hardware to control their own deployment, and to researchers who want to inspect or adapt the model. Hosted access still has provider terms, data handling rules and usage costs.
Our take
The release matters because the model is no longer limited to a hosted preview: developers can access the weights under Mistral's license and choose how to deploy them. The lower listed API rate also makes it worth evaluating for coding and agent workloads.
The important distinction is between availability and suitability. A million-token context, a strong security benchmark and downloadable weights do not guarantee a good result on your repository. Start with a small set of real tasks, track cost and review effort, and check the license and hardware requirements before committing to self-hosting.
Sources
- Mistral, Large 4 model card and license, checked October 29, 2026
- Mistral, model and API pricing, checked October 29, 2026
- Hugging Face, Mistral Large 4 weights and deployment, October 29, 2026
- Mistral, original public preview announcement, October 6, 2026
Spotted an error in this article? Report it · Our corrections policy
