What happened
Mistral’s announcement says the new weights are for Mistral Large 4, also called Le Chonk, a natively multimodal model with one trillion parameters. The model had entered public preview earlier in the month. The company says it used feedback from developers, early-access customers, and partners while continuing reinforcement learning for three weeks; it describes the resulting version as improved in coding, defensive cybersecurity, and reasoning efficiency. The supplied source calls the weights released, but does not specify a download URL, license terms, hardware requirements, or a price.
For coding, Mistral cites Cognition’s private FrontierCode 2.0 benchmark, which it describes as measuring end-to-end code quality and whether a repository maintainer would merge an agent’s change. Large 4 scored 19.7%, ahead of the listed comparison models in the reproduced chart. Mistral also cites the Artificial Analysis Coding Agent Index: it reports a 48% score, effectively tied with GLM-5.3, while using 47% fewer tokens per task on average. The announcement says Large 4 improved by 13 percentage points over its public-preview result on Terminal-Bench 4.0, reaching 41%, and reduced output tokens by 30%.
The release adds three reasoning modes to let developers trade off effort and cost: Low for simple tasks, replacing the former “None” mode; Medium for everyday tasks; and High, the default, for difficult tasks requiring precision and thorough reasoning. Mistral also cites blind pairwise evaluations by Surge Code, where it says Large 4 was preferred to several named open-weight models and ranked behind Claude Opus 5.5 among evaluated models.
For defensive cybersecurity, Mistral reports a 51% score on Artificial Analysis Cyber Index, described as testing vulnerability discovery, reproduction, validation, and patching without breaking existing functionality. It also reports an average 96.6% refusal rate for malicious cyber requests across prompts from JailbreakBench, StrongREJECT, and AgentHarm. The company says the model distinguishes defensive from malicious requests. Booz Allen is quoted describing internal source-code security testing and deployment customization; this is a customer statement, not an independently detailed evaluation in the supplied material.
What the archive adds
The archive has a prior note on Mistral Large 4’s open-weight release, whose thesis says the earlier release included coding-agent results. This announcement is a new fact on that known story: it describes the improved weights and additional benchmark claims, rather than a wholly new model event. The supplied earlier Telegram posts also describe the public preview, but do not establish the separate claim that the model built an orbital tracker.
Trend signal
This supports, but does not prove, agent economics moves to completed work: Mistral emphasizes token use per coding task alongside a score intended to reflect end-to-end code quality. That framing is more useful than token price alone, but the supplied material does not provide independent verification or actual production costs per successfully completed task.
Why now
The timing follows Large 4’s public preview earlier in the same month. Mistral says it spent three further weeks scaling reinforcement learning after gathering feedback from developers, early-access customers, and partners. The release therefore presents the new weights as an updated artifact, not simply an announcement of the preview version. The source does not identify a release schedule beyond that sequence or explain how feedback was selected for training.
Why it matters
For a team considering an open-weight coding model, the announcement suggests a test plan rather than a deployment conclusion. Compare the model on the team’s own repositories and task mix, record whether changes pass review and tests, and track total tokens and latency per accepted change. The reported 47% lower average token use is potentially relevant to inference economics, but it does not by itself establish lower total cost: serving infrastructure, retries, review effort, and task success all matter.
The reasoning modes are a concrete control surface to evaluate. Teams can try Low for routine edits, Medium for everyday work, and High for difficult changes, then compare quality and resource use. The cyber results may interest defenders assessing a model for vulnerability triage or patching, but the reported refusal and Cyber Index figures do not substitute for access controls, sandboxing, or security review.
Limits and uncertainty
The claims in the announcement are from Mistral and named benchmark summaries; the supplied material does not include benchmark protocols, full result datasets, independent replication, or the specific model artifact’s license and download details. It does not show how results vary across languages, repository types, hardware, or production workloads. The source does not establish that Large 4 created or verified the orbital-tracking app mentioned in the supplied social post. It also provides no evidence that any benchmark result predicts performance on a particular team’s tasks.
Skill to build now
Create a small evaluation sheet for one representative coding workflow this week. Run the same bounded tasks through Low, Medium, and High reasoning modes, then record accepted changes, test outcomes, retries, latency, and total input and output tokens. Keep defensive-security work in a sandboxed environment with review before applying patches. This turns the announced token-efficiency claim into a comparison against the team’s own completion and verification costs.
Reader takeaway
Mistral has published a new set of Large 4 weights and reports stronger coding results with lower token use on selected benchmarks. Treat the numbers as a reason to run a controlled evaluation, not as proof of cheaper or more reliable production work. The orbital-app claim is not substantiated by the supplied primary source.