Mistral released downloadable weights for an improved Mistral Large 4, nicknamed Le Chonk, on October 29, 2026. The release fulfills the weight-release promise made with the October 6 public preview and adds further coding and reasoning improvements. For businesses, the decision now includes self-managed deployment costs and the model’s commercial licensing conditions.
October 6 was the preview; October 29 delivered weights
Mistral’s October 6 announcement launched a public API preview through Mistral Studio and promised weights by month-end. It was an announcement and hosted-access event, not the downloadable release. The October 29 announcement explicitly releases an improved checkpoint after further reinforcement learning and feedback. API availability and weight availability should therefore not share a single launch date.
The first-party repository is mistralai/Mistral-Large-4-1T-A52B. Its model card identifies the hosted API model as mistral-large-4 and documents self-serving options. Downloading the checkpoint is a separate choice from calling that endpoint; it does not bring the provider’s managed service along with it.
The upgrade changes what teams should evaluate
Mistral reports that Terminal-Bench 4.0 performance reached 41%, around 13 percentage points above preview, while output-token use fell 30%. It also introduces low, medium, and high reasoning modes, with high the default and low replacing the previous none mode. These are publisher-reported results, not Nerova measurements or a guarantee of equivalent gains on a customer repository.
Our operational reading: a preview evaluation is no longer sufficient evidence for adopting the released checkpoint. Re-run the same repository tasks with fixed tool permissions and acceptance tests. Compare completed, accepted changes and total execution time—not just tokens. Start with medium reasoning, then check whether high actually resolves failures worth its additional work.
The license is not an unrestricted MIT grant
The actual LICENSE file names the Mistral Ecosystem Open-Weight License, a modified MIT license. It permits use, modification, further training, and redistribution subject to conditions, including preservation of notices. Crucially, if a licensee or affiliate operates a Model-as-a-Service business, commercial use of the model or derivatives requires a separate Mistral agreement, with stated exceptions.
The definition concerns third-party access for inference or post-training with meaningful control over inputs, parameters, or training data. It excludes capabilities embedded solely within specific end-user features or harnesses, and mere request relaying. Internal use and access through Mistral’s official products are exceptions to the separate-license requirement. Do not collapse this into “all commercial use is forbidden” or “every hosted application is exempt.”
For procurement, the practical task is to describe what customers can control and which affiliated businesses operate model services, then obtain legal review. Open weights describe access to the artifact; they do not erase contractual boundaries.
Self-hosting is real, but not lightweight
The model card describes roughly one trillion total parameters, with 49 billion active per token, or 52 billion including embeddings and output layers. It accepts text and images and produces text, with a stated one-million-token context. Mistral provides vLLM and SGLang recipes; an NVFP4 checkpoint targets Blackwell GPUs. The repository’s files view lists approximately 1.06 TB. These are deployment specifications, not a measured hardware-sizing result.
Our recommendation is to request a capacity test at the intended context length and concurrency before committing infrastructure. The active-parameter figure is not a substitute for that test. Evaluate the complete agent, including retrieval, tool execution, recovery, and operator review.
The public release is not every cybersecurity variant
Mistral separately describes a specialized cybersecurity variant available in preview to vetted enterprise design partners under contractual safeguards. Its public downloadable checkpoint should not be presented as unrestricted access to that variant. The company also acknowledges potential misuse and continuing guardrail changes.
For a first business evaluation, we would restrict a coding or defensive-security agent to an isolated repository and require approval before merging patches. The October 29 release makes independent operation possible; it does not establish that an agent is safe to act unattended.