Mistral Large 4's Coding Lead Tests Open-Weight Ambitions

Mistral released the weights for Large 4 on Hugging Face on October 29, giving developers access to a model the company says narrowly leads an open-weight rival on a coding test. The release follows a preview roughly three weeks earlier. Mistral says reinforcement learning and testing with customers and partners improved the model between preview and release.

The headline comparison is close: on Mistral's Terminal-Bench 4.0 figures, Large 4, nicknamed Le Chonk, scores 41%, against 40% for GLM-5.3. Mistral also says Large 4 uses 47% fewer tokens per task on average than GLM-5.3. Those claims describe two different potential advantages. One concerns how often an agent completes tasks in a test; the other concerns how much text it generates while doing so. A developer choosing a model needs to know whether both hold for the work they intend to run.

The one-percentage-point lead gives Mistral a defensible benchmark claim on its own figures. It gives GLM-5.3 a strong showing on the same test. Neither result settles how the models compare across repositories, tools and operating constraints outside that test. The token claim could be consequential even if the completion scores remain close, provided it holds on comparable tasks and under comparable settings.

A Bigger Model, and a Bigger Operational Bet

Mistral's route here helps explain why this release carries more weight than a new score on a chart. The company moved from smaller open models to Large 3 in late 2025, with Large 4 aimed at coding, reasoning, image analysis and business automation. That range gives the company several markets to pursue, but the coding comparison supplies one of the clearest public tests of its ambitions.

In a Hugging Face discussion with Mistral's science team, Pierre Stock described Large 3 as roughly 600 billion parameters and Large 4 as about one trillion. He also described Large 4 as the first launch trained using a newly online cluster of roughly 4,000 chips on Mistral's own compute. Stock said operating a large GPU cluster brought infrastructure challenges alongside the model-training work.

That account changes the question behind the benchmark. Mistral has invested in the capacity to train a substantially larger model and now has a release against which developers can assess the result. The extra scale may support more capabilities, but the parameter count alone cannot tell an organization whether Large 4 will finish its coding jobs more reliably or cheaply. The 41% score, the token-use claim and the work of running the model each answer a different part of that question.

Mistral's case for open weights is practical as well as competitive. In a press briefing attended by The New Stack, Stock argued that access to weights lets an organization run a model on its own infrastructure or infrastructure it trusts, reducing dependence on a provider that could withdraw access. For a team building a coding agent into its workflow, that is a concrete form of control: it can consider a deployment path beyond a single provider's hosted endpoint.

Control still has an infrastructure bill. Stock's description of Mistral's own training cluster shows how much machinery the company has put behind Large 4. That figure describes training, not what every user must assemble to run the released model. It does, however, sharpen a planning question for would-be adopters: what resources would their chosen deployment require, and who would operate them? Weight access creates an option; it does not answer that operational question for them. Nor does the term open weight by itself specify what permissions a license grants. Teams planning to modify or redistribute a deployment should check the applicable terms rather than infer them from the release label.

What the GLM-5.3 Comparison Can Tell You

GLM-5.3 is the useful rival here because Mistral supplies numbers for both models on Terminal-Bench 4.0. At 41% and 40%, their reported task-completion results are close. A team whose agents handle unfamiliar repositories could start with that comparison, then ask whether its own tasks produce a similar result. A team paying for every generated token might put more weight on Mistral's separate efficiency claim. Neither priority is irrational; they describe different costs of getting code work done.

Stock said in the Hugging Face discussion that GLM-5.3 received the same score as Mistral's model on an Artificial Analysis coding index. The discussion also describes an efficiency measure with a length penalty. That account should not be collapsed into the Terminal-Bench comparison: an index and a component test need not produce the same ranking, and a length-sensitive measure asks a different question from task completion alone.

Token use has a plausible operational payoff. If two agents finish comparable work at similar rates and one generates fewer tokens, a team may spend less on token-priced inference or free capacity for additional tasks. Fewer tokens alone cannot establish faster responses or a lower total deployment cost; model-serving costs and the time spent generating or checking a result also enter the decision. Stock's claim of 47% fewer tokens is therefore a useful question to test, rather than a price quotation.

The comparison also clarifies what a narrow lead does not demand from developers. They do not have to treat a one-point result as a mandate to migrate an existing agent. They can test both models on the same tasks, with the same tools and completion criteria, and record token use alongside successful outcomes. For a maintainer evaluating automated changes, the useful unit is a completed task that still survives review, not a leaderboard position detached from the project's workflow.

Large 4's release gives teams an additional open-weight candidate and gives Mistral a chance to show what its move to larger models and owned compute bought. Its claimed Terminal-Bench lead is narrow; its claimed token saving, if it travels beyond the company's test setup, could make the choice more interesting. The developer's next question is less glamorous than Le Chonk's nickname: on the jobs you actually need finished, what does each successful run cost you to produce and check?