What was released
On October 29, Mistral AI released the weights of Mistral Large 4, nicknamed Le Chonk, three weeks after a public preview. According to Hugging Face’s launch post, it is a multimodal, multilingual mixture-of-experts model with about 1 trillion total parameters, about 49 billion active per token, and a 1 million token context window. It takes text and images in and produces text.
Two official variants are available, FP8 and NVFP4, plus community GGUF builds. The Next Web reports a quantized version runs on four Nvidia B200 or B300 GPUs, so this is not a laptop model.
trilUnder the hood
Hugging Face describes a 72-layer decoder with a hidden size of 6,144 and 64 attention heads. Attention combines Multi-head Latent Attention with a sparse indexer that selects up to 2,048 key positions per query. In the mixture-of-experts layers, a router picks four of 128 experts per token, with one shared expert always active.
Put simply, the model is huge on paper but uses only a slice of itself for each token, which keeps inference cost down.
yers,Coding gains and efficiency
Mistral says three extra weeks of reinforcement learning, guided by customer feedback, improved the model. The Next Web reports Terminal-Bench 4.0, a test of agentic coding in a terminal, rose from 28% at preview to 41%, just ahead of GLM-5.3 at 40%. On Artificial Analysis’ Coding Agent Index the model rivals GLM-5.3 while using 47% fewer tokens per task on average.
Mistral also added low and medium reasoning modes alongside high and made high mode 50% more efficient, giving buyers a way to trade cost against depth. Benchmark results are Mistral’s and third-party index figures as reported; treat them as vendor-reported.
ArtifThe license catch
The weights ship under the Mistral Ecosystem Open-Weight License, which Hugging Face describes as an MIT derivative with limits on model-as-a-service offerings. Companies can run it internally, but anyone selling hosted inference to third parties needs a separate commercial license. So open here means open to use, not open to resell as a service.
d parPractical takeaways
- Check the license before building a hosted product on the weights.
- Plan hardware: the quantized build needs four top-end GPUs.
- Compare reasoning modes on your own tasks; the efficiency claims are vendor-reported.
Bottom line
Mistral Large 4 is one of the largest open-weight models available and a strong play for European, self-hosted AI, with a license that keeps cloud resellers on a leash.