Mistral open-weights Le Chonk, a 1-trillion-parameter model that leads open coding benchmarks

Mistral released the full weights of Mistral Large 4 on Thursday, three weeks after the 1-trillion-parameter multimodal model entered public preview. The Paris lab published the release notes under the internal codename Le Chonk, and the pitch is unusually specific: best open-weight model for coding, a defensive-security story backed by named vendors, and a token-efficiency claim that undercuts the usual “bigger is better” framing.

The headline number is the one to argue about. On Cognition’s private FrontierCode 2.0 benchmark, which measures whether a repository maintainer would actually merge an agent’s change, Mistral Large 4 scores 19.7%. That is a low absolute number, and Mistral frames it as a lead. Read it as a sign of how early agentic coding still is, not as a victory lap. The more interesting figures sit next to it: on the Artificial Analysis Coding Agent Index the model scores 48%, effectively tied with GLM-5.3 while using 47% fewer tokens per task. On Terminal-Bench 4.0 it jumped 13 percentage points to 41% over the preview, with 30% fewer output tokens.

That efficiency claim is the actual news. A trillion-parameter model that matches a competitor on agent quality while burning roughly half the tokens changes the unit economics of running coding agents at scale. Token count is the line item that quietly decides whether an agent loop is affordable in production. Mistral is selling cost-per-task, not cost-per-parameter.

Three reasoning modes, one tradeoff

Mistral now exposes three reasoning modes: Low for simple tasks (replacing the old “None” setting), Medium for everyday work, and High as the default for hard cases. This is a sensible concession to reality. Most agent traffic is not hard, and paying High-mode reasoning on a routine refactor is waste. The risk is that developers leave everything on the default and never tune it, which would erase the efficiency gain the release is built around. Mistral’s own benchmarks assume the modes are used deliberately.

The human-preference data adds a useful check on the benchmark theater. In Surge Code’s blind pairwise evaluations, graders preferred Mistral Large 4 over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, ranking it behind only Claude Opus 5.5 among the models tested. That puts a closed frontier model ahead of it, which is the honest framing: Mistral leads the open field and trails the closed frontier. It says so, more or less, in its own materials.

The cybersecurity bet is the real strategy

The coding numbers will get the attention. The defensive-security positioning is where Mistral is making a structural bet. The model scores 51% on the Artificial Analysis Cyber Index, which tests whether a model can find a vulnerability in a codebase, reproduce it, validate it, and patch it without breaking existing functionality. Mistral stresses that it handles these legitimate defensive tasks “without unnecessary refusals,” and reports an average refusal rate of 96.6% on malicious cyber prompts across JailbreakBench, StrongREJECT, and AgentHarm.

That pairing, capable on defense and refusing on offense, is the whole sales pitch to governments and regulated enterprises. The vendor quotes Mistral assembled read like a procurement checklist: Booz Allen Hamilton’s Brad Medairy cites data-residency requirements and a “compelling cost curve”; Cloudflare’s Stephanie Cohen argues that “a defense you don’t control isn’t defense”; CrowdStrike’s Dr. Bartley Richardson frames it around remediation speed; ENISA’s Laura Heuvinck confirms the EU agency is testing the model to find vulnerabilities and audit code. Cisco, Fortinet, Elastic, Zscaler, Rubrik, Orange Cyberdefense, and NVIDIA’s Kari Briski round out the list.

Mistral is not really selling a model here. It is selling a procurement argument: own the weights, host them yourself, and stop routing your security data through someone else’s cloud.

That is the sovereign-AI category Mistral says it is defining, and the release leans on it hard. The model was “trained from scratch on European infrastructure with clear lineage from inception to release.” For a European government or a bank with residency rules, that lineage is the product. The capability is table stakes.

What the release admits

Buried near the end, the post concedes something worth flagging. Mistral says it monitored live traffic and, while it saw the model refuse black-box attacks against live targets, “we see a potential for uplift to conduct real attacks. We are modifying our guardrails accordingly.” The source text is truncated at that point, so the full scope of the guardrail change is unclear. But the admission matters: releasing weights for a model this capable at vulnerability work carries a dual-use risk that the benchmark slide does not capture. Mistral is naming it rather than hiding it, which is the right move and also an acknowledgment that open weights cannot be un-released once they are out.

The compute story is quieter but present. A 1T-parameter model trained on European infrastructure implies a serious cluster, and NVIDIA’s inclusion in the partner quotes suggests accelerated hardware is doing the training and inference. Mistral does not disclose the chip count or the training budget here. Those numbers would tell us how much capital it takes to sit at the open frontier, and their absence is the biggest gap in the release.

What it means for builders

For teams running coding agents, the practical question is whether the token-efficiency claim holds outside Mistral’s own harness. The 47%-fewer-tokens figure comes from Artificial Analysis, which is at least a third party, but agent frameworks vary wildly in how they spend tokens. A team should measure cost-per-merged-PR on its own repo before trusting the headline.

For the open-weight field, the release raises the bar for GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash without settling anything. The gap to Claude Opus 5.5 on human preference is the number to watch next. If Mistral closes it in the next reinforcement-learning cycle, the “open trails closed” assumption gets harder to defend. If it stalls, the frontier stays where it is.

The clearest signal is the customer list. Booz Allen, CrowdStrike, Cisco, Cloudflare, and a European regulator all put their names on a model release. That is not a developer-community launch. That is a government-and-enterprise sales motion wearing open-weights clothing, and it tells you where Mistral thinks the money is.