Mistral Large 4 weights released October 29, 2026; public preview October 6, 2026
Mistral Releases Large 4 Weights Under a Modified MIT License, With Its Cyber Variant Held Back for Vetted Partners
Mistral released the weights of Mistral Large 4, nicknamed Le Chonk, on October 29, making the roughly 1-trillion-parameter mixture-of-experts model downloadable from Hugging Face for self-hosting three weeks after its October 6 public preview, following what the company describes as three additional weeks of reinforcement learning. The model has roughly 49 to 52 billion active parameters per token depending on how embeddings are counted, a 1-million-token context window, and text and image input. It is open-weight, not open source in the unrestricted sense: the Mistral Ecosystem Open-Weight License, a modified MIT license, allows use, modification, fine-tuning, and commercial products, with one condition that businesses operating a Model-as-a-Service offering must obtain a separate commercial license first. Through Mistral own API the model costs 0.70 dollars per million input tokens and 2.10 dollars per million output tokens, and it is also carried through Microsoft Foundry, described as the first cloud to offer it, and as an NVIDIA NIM container. Mistral pitches the model at cyber defenders, reporting a 51 percent score on the Artificial Analysis Cyber Index, 48 percent on the Coding Agent Index, 41 percent on Terminal-Bench 4.0, and a 96.6 percent average refusal rate on malicious cyber requests across the cyber prompts of three safety benchmarks, figures the company and its partners report rather than independent tests. The release is unusually direct about dual use: Mistral says it monitored cyber-related use during the preview, saw a potential for uplift to conduct real attacks, and is tightening guardrails, while a specialized variant built for penetration testing and red-teaming goes only to vetted partners such as Orange Cyberdefense under contract terms restricting testing to owned or explicitly authorized systems. Open weights cannot be recalled once downloaded, which is both the pitch to governments and banks that cannot send code or logs to an outside API and the standing risk the announcement itself acknowledges.
Executives
The strategic choice in this release is sovereignty against control: a downloadable frontier-class model can run entirely inside your own boundary, which solves data residency and vendor cutoff concerns, and it can never be patched, recalled, or remotely shut down by its maker. Both halves belong in the decision record. Note also that the license is not plain MIT: if your business model involves hosting models for others, the separate commercial license is a procurement step, not a detail.
CIO / CISO
Treat an open-weight deployment as a system you fully own under the NIST AI RMF: your organization becomes the party responsible for guardrails, monitoring, updates, and incident response, because no vendor control plane sits behind it. Record the exact checkpoint and license terms in your AI inventory, plan how you will track guardrail research against a model whose refusals a determined operator can attempt to fine-tune away, and keep the restricted cyber variant distinction in mind when vendors offer penetration testing features built on this family.
Engineers
Plan capacity from the checkpoint sizes, not the parameter headline: full precision weights run to roughly a terabyte across dozens of files, with quantized variants and an 8-GPU node, a 4-GPU Blackwell node, or a high-memory Apple Silicon system among the demonstrated hosting shapes, and Mistral itself states no memory requirement. Pin the exact Hugging Face revision you deploy, verify its hashes, and serve it behind your own authentication and logging from day one, since the weights ship without any of that.
Everyone (non-IT)
A capable AI model that anyone can download and run on their own computers is now a normal event, not an exception. For most people nothing changes directly, but it is the reason your employer may soon offer an internal AI tool that runs entirely on company systems, with your prompts never leaving the building.