Mistral Large 4 open weights: coding-lead claims still need third-party replication

Mistral moves the story from another flagship preview to downloadable weights. For buyers and security teams, the real variable is not the Sovereign AI slogan but whether a self-hosted 1T multimodal model reproduces the charted leads on agentic coding and defensive cyber workflows in independent environments.

The falsifiable reading is narrow: open weights change the deployment and audit choice set; FrontierCode 19.7%, Coding Agent Index 48%, and Cyber Index 51% remain company- and partner-reported rankings until third parties fully replicate them—not a closed industry verdict.

Hard facts in the post: weights cover the improved Large 4; the model is described as trillion-parameter and natively multimodal; post-preview RL focused on coding and defensive cybersecurity; three reasoning modes trade cost for performance; Surge Code blind prefs are said to beat several open rivals and trail only Claude Opus 5.5; named blurbs cite Booz Allen, Cisco, Cloudflare, CrowdStrike, Elastic, ENISA, Fortinet, NVIDIA, Orange Cyberdefense, Rubrik, and Zscaler.

Mistral also writes that live traffic showed refusals of black-box attacks on live targets, yet it sees potential uplift for real attacks and is modifying guardrails; the specialized cyber variant is limited to vetted partners under contractual safeguards on owned or explicitly authorized systems.

The evidence structure is a lab blog plus named partner quotes and selected benchmark charts—not peer review or a fully independent leaderboard a reader can reproduce on their cluster today. FrontierCode is labeled Cognition’s private benchmark; some Terminal-Bench comparisons include a self-reported preview point. Human prefs come from Surge Code; sample and prompt distributions are not fully unpacked in the extract.

The steelman: open weights let red teams and buyers audit architecture, land private data, and customize under residency constraints—the common theme across the security executive quotes. That steelman does not automatically cover “the percentage-point lead on the chart holds on your GPUs.”

Strongest steelman for Mistral: while closed APIs dominate agentic coding, releasing self-hostable weights that emphasize refusing malicious cyber requests turns “European lineage + customizable” from marketing into an executable deployment option; three reasoning modes also answer token cost directly.

The limit is verification: private benches, partner endorsements, and the company’s own attack-uplift warning all require splitting “open” from “proven stronger/safer.” License terms, hardware floors, and a full model card are not replaced by the blog extract.

For security ops, the next months’ observables are independent public replications of the Cyber Index / coding-agent scores, guardrail update cadence, and whether the specialized red-team variant stays inside tight contracts. For sovereign buyers, the test is latency, cost, and compliance audits on owned infrastructure—not another endorsement list.

For the open ecosystem, 1T-class weights lower the narrative that only closed models can do agentic coding, while enlarging misuse and distillation surfaces—the company’s own sentence on attack uplift is the footnote on that dual edge.

Over the next six months, if only Mistral and partner charts exist without citable third-party full replications and a stable guardrail changelog, Large 4 open weights should be read as a sovereign deployable supply shock and an auditable coding/cyber option—not a final finding that open weights systematically beat closed models on agentic coding. Only checkable independent leaderboards and public red-team replications would justify revising capability-diffusion and safety baselines upward.

【No.2】