Le Chonk’s open weights arrive with hosting terms and new coding scores

The small print is part of the release

Mistral released the weights of its improved Large 4 model, Le Chonk, on October 29. The downloadable release follows a public preview earlier in October. For developers, the most immediately useful distinction is not just between the old and new benchmark scores. It is between access to the model and the terms under which a business can offer access to others.

The published model licence is named the Mistral Ecosystem Open-Weight License and describes itself as a modified MIT licence. Its text contains a separate commercial-licensing condition for a defined Model as a Service business, along with definitions and exemptions. It is not accurate to replace those terms with the shorthand that open weights mean unrestricted commercial hosting. The model repository provides the actual document a builder should assess for its intended use.

That makes the release relevant to several different business models. A company embedding a model inside an application and a company selling general access to the model should not assume they occupy the same licensing category. This is an interpretation of the product decision, not legal advice or a determination that a particular deployment qualifies for an exemption. The exact arrangement matters more than the label open.

A preview-to-release comparison, with attribution

Mistral describes Large 4 as a one-trillion-parameter multimodal model and says additional training after the preview improved coding and defensive-security performance. Its release announcement reports a Terminal-Bench 4.0 score rising from 28% in the preview to 41% for the new version. The difference is 13 percentage points, not a claim that every coding task is now completed 13% faster.

The company also reports 19.7% on Cognition's private FrontierCode 2.0 benchmark. The release describes that benchmark in terms of end-to-end code quality and whether a repository maintainer would merge a change. This is a different measurement from merely producing code that looks plausible. The scores remain vendor-reported; Foundra has not reproduced either evaluation.

The two figures should not be placed on a single generic scale of intelligence. Different tasks, scoring rules and evaluation conditions can make numerically similar results mean different things. A reader can use the company comparisons to identify what changed in the release without converting them into an independently verified ranking of products.

This is also why a score well below 100% is informative rather than embarrassing by definition. Its meaning depends on the task set and what counts as success. A demanding assessment can reveal useful differences among models while still leaving substantial work unfinished. The release is a model-development milestone, not proof that software engineering has become an unattended process.

Three reasoning modes change a product setting

Large 4 offers low, medium and high reasoning modes. Mistral presents them as choices for different task demands. They give application builders a control over how the model is used, but the names do not themselves define service quality or the total cost of completing work.

An application can measure that trade-off without claiming that high always wins. The comparison needs a fixed task and a defined result: for instance, an accepted change rather than an answer emitted. The resources consumed before that result can include retries and review as well as inference. That is a measurement distinction, not an evaluation outcome we have observed for this release.

The commercial implication is that model cost and application cost are related but different. A cheaper response that is not usable may not be a cheaper service. A more costly inference step might be justified by a better completed result, but only if the full workflow supports that conclusion. Neither proposition follows automatically from the settings menu.

Weights, capability and permission to act

The release expands a builder's deployment options. It does not expand an assistant's authority by default. Downloading a model, evaluating its output and authorising it to change a live system remain separate decisions.

Mistral's emphasis on defensive cybersecurity makes that boundary especially important. A product can use a model to assist an authorised investigation without giving it unrestricted access to targets or production credentials. The boundary belongs to the application and its operator, not just to the model's reported benchmark performance.

Large 4's new weights, revised results and published terms are therefore three different parts of the same product news. The weights offer an infrastructure option; the results describe claimed improvements; the licence defines conditions that a business must examine. Treating all three as one undifferentiated promise of open AI would make the release easier to summarise and harder to use.