Mistral makes the updated Large 4 weights available after its public preview

Mistral released the updated Large 4 weights following the hosted preview.

Mistral followed its October 6 public preview of Large 4 with a downloadable checkpoint at the end of October. The company says it ran roughly three additional weeks of reinforcement learning after the preview. The resulting release combines two changes for developers. The model now has a different set of trained weights, and organizations can evaluate those weights on infrastructure they control. [Original preview](https://mistral.ai/news/mistral-large-4/) · [Weights announcement](https://mistral.ai/news/mistral-large-4-oss/) · [Checkpoint](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B)

What the published results say

Mistral reports that Large 4 scored 19.7% on Cognition's private FrontierCode 2.0 benchmark, which assesses the quality of an agent's proposed changes. It reports a 48% score on the Artificial Analysis Coding Agent Index and says the model used 47% fewer tokens per task than GLM-5.3 in that comparison. These results describe different tests. Neither token usage nor the private benchmark alone establishes what a team will pay to complete a real repository task. [Mistral's evaluation summary](https://mistral.ai/news/mistral-large-4-oss/)

One practical comparison would hold the repository snapshot, tools and acceptance tests constant while changing the checkpoint. It could record whether the agent fixed the stated problem, whether a maintainer accepted the patch and how much review was needed. Logging time and inference cost together would expose cases where a shorter response requires more work after the agent stops.

What changes for self-hosting

Mistral describes Large 4 as a natively multimodal model with around one trillion parameters overall. The repository's name indicates 52 billion active parameters. The distinction matters because self-hosting capacity depends in part on the size of the distributed weights, while per-token compute depends on how the runtime uses the model. Capacity planning also needs to account for request length, concurrency and any processing of visual inputs. [Mistral announcement](https://mistral.ai/news/mistral-large-4-oss/) · [Checkpoint specifications](https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B)

The weights ship under the Mistral Ecosystem Open-Weight license. Teams can evaluate the published checkpoint and inspect the license for their use. Mistral also describes a separately vetted preview of a specialized cybersecurity variant. That arrangement has its own scope and terms. [Release and preview access](https://mistral.ai/news/mistral-large-4-oss/)