Large 4’s coding pitch: equal scores, fewer tokens

What changed after the preview

Mistral positions Large 4, also called Le Chonk, as its flagship open-weight model for coding and cybersecurity. The release brings improved weights following a public preview. In its announcement, Mistral attributes the gains in coding and terminal workflows to three additional weeks of reinforcement learning.

Mistral describes Large 4 as a hybrid instruct-and-reasoning mixture-of-experts model. It supports multimodal input, rather than being presented as a coding-only model.

The practical pitch has two parts. Mistral cites coding evaluations that place Large 4 alongside or ahead of other evaluated open-weight models. It also says organizations can inspect, customize, and deploy the weights on their own infrastructure without depending on Mistral to serve the model through an API.

Those are separate reasons to evaluate it. The coding results concern performance under particular evaluations. The downloadable weights concern control over customization and deployment.

For engineers, the useful reading of this launch is therefore narrower than Mistral’s broad claim of open-weight coding leadership: improved weights, reported reductions in token use, selectable reasoning effort, and multiple deployment paths.

What the coding results establish

Mistral cites Artificial Analysis’ Coding Agent Index, which gives Large 4 a score of 48%, tied with GLM-5.3. Artificial Analysis separately reports 47% fewer tokens per task on average.

Keep those measurements distinct. The score establishes the reported tie on that index. The token figure adds an efficiency claim.

Mistral also cites three other coding evaluations:

• Surge Code: In blind pairwise evaluations, Large 4 was preferred over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash. It ranked behind only Claude Opus 5.5 among the models evaluated.
• Terminal-Bench 4.0: Mistral reports a score of 41%, an improvement of 13 percentage points over the public preview, alongside a 30% reduction in output tokens. Mistral says the result takes the lead among open-weight models.
• Cognition’s private FrontierCode 2.0: Mistral reports a leading score of 19.7%. The benchmark assesses end-to-end code quality and whether a repository maintainer would merge an agent’s change.

These results support different comparisons. Surge Code reports preference among its evaluated models. Terminal-Bench supplies a before-and-after comparison with the preview. FrontierCode targets the quality and mergeability of a completed change, but is a private benchmark.

Large 4 can tie GLM-5.3 on one index and be preferred over it in a separate evaluation without either result contradicting the other.

Likewise, the two token reductions are not interchangeable. Artificial Analysis reports tokens per task. The Terminal-Bench comparison specifically reports output tokens relative to the preview. They should not be combined into a single efficiency estimate.

Reasoning effort and the API bill

Mistral exposes three reasoning modes, with High as the default:

• Low: Intended for simple tasks and replaces the previous “None” reasoning mode.
• Medium: Intended for everyday tasks.
• High: Intended for the hardest use cases requiring high precision and thorough reasoning.

Mistral presents these settings as a way to balance performance against cost. That makes the reasoning setting part of the configuration to record during an evaluation, rather than a detail to leave implicit. A test using the default is a test using High.

Mistral’s listed API prices are:

• Input: $0.70 per million tokens.
• Cached input: $0.07 per million tokens.
• Output: $2.10 per million tokens.

The reported token reductions make the pricing relevant, but they do not by themselves establish a dollar saving for a particular workload. The price schedule distinguishes input, cached input, and output. Artificial Analysis’ reported reduction is expressed as tokens per task, while the Terminal-Bench reduction is specifically about output tokens.

For an engineering evaluation, retain those categories rather than collapsing everything into total tokens.

Low, Medium, and High are controls to test, not evidence that a particular setting will preserve the headline coding score at a lower bill.

Open weights, with a service-license boundary

Mistral says Large 4 is available through Mistral Studio, the Mistral API, OpenRouter, and Vibe Code. For self-hosting, it directs engineers to weights on Hugging Face and deployment with vLLM or SGLang.

The announcement also lists NVIDIA-hosted endpoints and Microsoft Foundry. Mistral says Foundry offers private deployment options for Azure developers. The supplied availability information does not specify consumer app tiers or regions.

The important distinction is between being able to obtain the weights and having unrestricted commercial rights for every business model. According to the release details, the Mistral Ecosystem Open-Weight license permits use, modification, fine-tuning, and commercialization. Commercial Model-as-a-Service operators need a separate license.

A team evaluating Large 4 for an internal application and a company planning to serve the model commercially should not assume that the same licensing conclusion covers both cases. The release explicitly identifies commercial Model-as-a-Service operation as a separate licensing case.

Mistral’s broader deployment claim is that organizations can inspect and improve the model without relying on its API. That is the operational choice the weights provide. It is not a reason to describe the license as unrestricted.

For teams comparing hosted and self-managed paths, keep the decision grounded in the available options: Mistral and other hosted access points, Azure private deployment through Foundry, or downloaded weights with the named serving frameworks. The announcement provides those routes, not a complete self-hosting configuration.

Cybersecurity capability and restricted access

Mistral makes a separate cybersecurity case for Large 4. It cites a 51% score on the Artificial Analysis Cyber Index, which tests vulnerability discovery, reproduction, validation, and patching without breaking existing functionality.

Mistral also reports a 96.6% average refusal rate on malicious cyber requests, describing it as the highest it measures among open models. That is Mistral’s measured refusal claim, not an unconditional statement about how the model will behave on every malicious request.

The announcement contains an important qualification. Mistral says monitoring of live traffic showed the model refusing black-box attacks against live targets, but also identified potential for the model to help conduct real attacks. The company says it is modifying its guardrails and will continue updating them.

Separately, Mistral is providing a specialized cybersecurity variant in preview to vetted enterprise design partners, including Orange Cyberdefense. It describes that variant as optimized for penetration testing, vulnerability assessment, and red-teaming. Access requires vetting and contractual safeguards, with testing restricted to owned or explicitly authorized systems in controlled environments.

Do not read the general Large 4 availability list as an access promise for that specialized variant. The announcement presents a broadly available model and a restricted enterprise preview as different access paths.

Mistral also says customers are using Mistral Forge to tailor Large 4 with proprietary security data, policies, detection logic, and threat environments. That customization claim is distinct from both the Cyber Index result and the specialized variant’s access conditions. Each needs to be evaluated on its own terms.

Who should evaluate it

Large 4 is worth evaluating for coding-agent teams that want to test Mistral’s reported combination of competitive task results and lower token use. The strongest starting evidence is the Artificial Analysis tie, supported by the separate Surge Code, Terminal-Bench, and FrontierCode results.

For teams already working with the preview, the most directly relevant launch claim is the Terminal-Bench improvement: 13 percentage points higher, with 30% fewer output tokens. That is the supplied comparison between preview and release weights. It does not establish the same change for every existing workload.

For teams seeking deployment control, the relevant offer is inspectable, customizable weights with named self-hosting frameworks. Review the Model-as-a-Service licensing exception before making that route part of a commercial serving plan.

For security teams, distinguish the general model’s reported Cyber Index performance from the specialized preview and its approval requirements. Mistral’s own account of potential attack uplift is also a reason to retain the guardrail qualification alongside the capability scores.

Try it

Start in Mistral Studio or with the Mistral API. For self-hosting, start with the weights on Hugging Face and the named deployment frameworks, vLLM or SGLang. The supplied sources include no runnable commands.

Sources

• https://mistral.ai/news/mistral-large-4-oss/