Le Chonk Weights Drop: Europe’s 1T Model Goes Open

Posted on by
Contents show
Le Chonk

Mistral has released the open weights of Mistral Large 4, the trillion-parameter flagship the company nicknamed Le Chonk, and the move resets expectations for what open models can do. and the Le Chonk release reads less like a model drop and more like a strategic shipment. The Le Chonk weights landed on Hugging Face on October 29 in FP8 and NVFP4 checkpoints under the Mistral Ecosystem Open-Weight license, three weeks after the public preview, and the model that shipped is measurably stronger than the one that previewed: Le Chonk now leads open-weight models on Cognition's FrontierCode 2.0, leads on Terminal-Bench 4.0, and posts a 51 percent score on the Artificial Analysis Cyber Index while refusing 96.6 percent of malicious cyber requests.

The Le Chonk combination is the point. A frontier-capable model that anyone can download, inspect, fine-tune, and self-host changes the calculus for two audiences at once: enterprises that want coding performance without sending code to an API, and security teams that need defensive capability without a provider's refusal layer deciding mid-incident what work is allowed.

This article unpacks what the Le Chonk weights release actually contains, how the model earned its coding and cybersecurity numbers, what the license permits and forbids, and what Le Chonk's trajectory says about the open-weight race heading into 2027.

What shipped and how it got stronger in three weeks

The Le Chonk preview model announced in early October was a 1 trillion-parameter, natively multimodal mixture-of-experts system with roughly 50 billion active parameters, trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters. The weights release is that model plus three more weeks of reinforcement learning, and the improvement curve is unusually steep for a late-stage model.

The Le Chonk headline gains concentrate where Mistral aimed them. On FrontierCode 2.0, Cognition's private benchmark that scores end-to-end code quality by asking whether a repository maintainer would merge the agent's change, Le Chonk leads all models at 19.7 percent. On Terminal-Bench 4.0 it improved by 13 percentage points over the preview to 41 percent, first among open-weight models, while emitting 30 percent fewer output tokens. On the Artificial Analysis Coding Agent Index it scores 48 percent, effectively tied with the leading open-weight competitor, while using 47 percent fewer tokens per task.

That Le Chonk efficiency figure deserves as much attention as the scores. Token count is the operating cost of agentic coding, and a model that matches the leader while burning half the tokens per task shifts the economics of running coding agents at scale. Mistral credits the gains to its reinforcement learning pipeline, which continues training on the model's own attempts with rising task difficulty, and reports the run was still in flight at release with no signs of saturation.

Three Le Chonk reasoning modes, low, medium, and high with high as default, let developers trade precision against cost per task, which matters for teams routing mixed workloads through one model.

Inside the Le Chonk mixture-of-experts design

The architecture choices behind Le Chonk explain both its capability and its hardware demands, and they are worth unpacking for anyone planning a deployment.

A trillion-parameter model like Le Chonk with roughly 50 billion active parameters is a mixture-of-experts design: the full parameter set contains dozens of specialist subnetworks, and a router activates only the fraction relevant to each token. The result is the capacity of an enormous model with the per-token compute cost of a much smaller one, which is how Le Chonk delivers frontier-class answers while keeping inference economical. The trade-off is memory: all trillion parameters must sit in accelerator memory even though only a slice computes, which is why self-hosting requires server-class hardware rather than a workstation.

Natively multimodal means Le Chonk had vision trained in from the start rather than bolted on. The model reasons across documents, charts, natural images, and dense scenes, and Mistral's visual grounding results are the rare place where it claims superiority over closed frontier models, surpassing GPT-6 Astra on the Dense 200 benchmark by a point. Applied demos show the grounding stack doing real work: inspecting gigapixel satellite imagery for disaster response, verifying mechanical parts in engineering drawings, and retrieving evidence from long PDFs.

The Le Chonk training provenance rounds out the design story. Everything ran on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, on the same training, customization, and reinforcement learning environment the company sells customers through Mistral Forge. The reinforcement learning stage generates roughly 33 billion tokens per day at current scale, about 16 billion of them trainable completions, with environments spanning single-turn chat, scientific problem solving, safety alignment, and long-horizon tool use composed into single runs.

The cybersecurity story is the real differentiator

The most distinctive Le Chonk claims in the release are defensive, and they resolve a tension that has defined AI security tooling since the first refusal messages appeared.

Le Chonk scores 51 percent on the Artificial Analysis Cyber Index across discovery, reproduction, and patching, which measures whether a model can discover vulnerabilities in a codebase, reproduce and validate them, and patch them without breaking existing functionality. On the index's reproduction-and-patch test, the preview model scored 82 percent, the highest of any model evaluated, closed or open. Several leading closed models score near zero on that same test, not because they lack capability but because they refuse the task, since reproducing a real vulnerability in live software looks identical to attacking it until a patch exists.

That refusal gap is the strategic opening. Defending software often begins with proving a flaw is real, and security teams operating closed models lose that capability at exactly the moments it matters most, mid-incident. An open-weights model like Le Chonk self-hosted under the organization's own policies has no provider-level refusal layer, and Mistral's guardrails are built into the model itself: it retains strong performance on legitimate defensive work while refusing 96.6 percent of malicious cyber requests drawn from JailbreakBench, StrongREJECT, and AgentHarm, the highest measured rate among open models.

The company is careful with the edge of this capability. A specialized variant optimized for penetration testing, vulnerability assessment, and red-teaming is available only to vetted enterprise design partners such as Orange Cyberdefense, under contractual safeguards restricting testing to owned or explicitly authorized systems. Live-traffic monitoring during the preview window observed the model refusing black-box attacks against live targets, while the vendor notes potential for uplift toward real attacks and says guardrails will keep updating.

Sovereignty is the business model, and the license encodes it

The word Mistral uses repeatedly for Le Chonk is sovereignty, and the release operationalizes it in layers rather than treating it as a slogan.

The Le Chonk foundation layer is the model itself: trained from scratch on European infrastructure, with what the company describes as clear lineage from inception to release, downloadable and inspectable. Organizations can run Le Chonk on their own hardware with vLLM or SGLang, customize it with proprietary data through Mistral Forge, and operate the resulting agents through Mistral Studio without depending on Mistral's API for continuity. A managed path exists too, through Microsoft Foundry with private deployment options for Azure developers, plus NVIDIA-hosted endpoints via an optimized NIM container.

The Le Chonk license is the interesting instrument. The Mistral Ecosystem Open-Weight license is a modified MIT: free to use, modify, fine-tune, and commercialize, with one condition, businesses operating a Model-as-a-Service offering must obtain a separate commercial license. That single Le Chonk carve-out is aimed at the reseller layer, not the user layer. A company building products on Le Chonk owes nothing; a company renting Le Chonk out as a service competes with Mistral and pays for the privilege. It is a licensing shape the industry will likely see more of, because it funds open releases without surrendering the API market.

The Le Chonk pricing backs the openness. At 70 cents per million input tokens, 7 cents for cached input, and $2.10 per million output tokens, Le Chonk undercuts the frontier closed tier by a wide margin while offering self-hosting as the zero-margin alternative.

What the human evaluations add

Benchmarks invite tuning skepticism, so the blind human results for Le Chonk carry weight. In Surge Code's blind pairwise evaluations, professional graders preferred Le Chonk's output over leading open-weight competitors including GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, ranking it behind only Claude Opus 5.5 among models tested.

The earlier preview evaluation added breadth: professional annotators rated its coding second of five models, ahead of Kimi K3 and two GLM variants, behind only Claude Opus 5. Internal expert annotators compared it against GLM-5.3 across coding, CAD, finance, mathematics, and physics, preferring it in CAD and STEM while finding it on par or close in finance and coding.

The enterprise verticals round out the picture. Le Chonk reaches 73.3 percent on Finch, the finance spreadsheet workload benchmark, second among open models to Kimi K3, exceeds GPT-6 Astra on third-party-evaluated legal and financial tasks according to Mistral, and posts 84.5 percent on BrowseComp for web browsing, ahead of several frontier peers. On agent security, the B3 benchmark attack-resistance score rose from 93.3 percent in preview to 97.1 percent, which Mistral describes as the top of the open-source frontier, meaning self-hosters get a safe agent baseline without additional tuning.

What the demos show in practice

Benchmarks measure capability; the release demos show what that capability looks like when pointed at concrete work, and three of them are worth describing.

The Le Chonk vulnerability triage demo is the defensive story operationalized: an agent reads incoming vulnerability reports, investigates the source code and product context, reproduces each flaw in an isolated sandbox, then writes and iterates on a patch until a canary exploit no longer triggers. That pipeline is the daily workload of every security team drowning in disclosure volume, and the model performs it end to end without a provider deciding whether each step is permissible.

The Le Chonk agentic web development demo is the coding story compressed into one task: with no human help, the model built a live 3D map of roughly 19,000 satellites and debris pieces orbiting Earth. It pulled real tracking data, wrote the application, used the running product in a browser to catch its own bugs and slowdowns, then fixed what an independent reviewer flagged until every requirement passed. The notable behavior is the self-testing loop, where the agent treats its own output as a running system to be probed rather than code to be shipped blind.

The quantum chemistry demo is the Le Chonk science story: the model wrote, from scratch, software that calculates where electrons sit in a molecule, checked its answers against a standard chemistry tool, and then wrapped the solver in an interactive 3D page where dragging a slider pulls a molecule apart and shows electrons regrouping as bonds break. Earlier benchmarks reported the model generating a full Hartree-Fock simulation in one shot, and state-of-the-art status among open-weight models on SciCode-Verified.

Together the demos argue something benchmarks cannot: that the capability transfers to unfamiliar, multi-step, tool-using work rather than only to the distributions the evaluations were drawn from.

The open-weight race context

The Le Chonk release lands in a market where open-weight models have been closing the frontier gap all year, and the competitive frame matters for reading the numbers.

The 48 percent Coding Agent Index score ties the leading open-weight model while winning on efficiency, and the cyber results are the clearest open-weight lead anywhere: no closed model serves equivalent defensive capability at any price, because the refusal layers that make closed models safe for general release also block the work defenders need. That asymmetry is structural, not incidental, and open weights are currently the only delivery mechanism that resolves it.

The capital behind the Le Chonk push is also new. The release is the first milestone on the roadmap funded by a 3 billion euro Series D, described as the largest equity round ever raised by a European technology company, with compute capacity expanding in Mistral's own datacenters. The training data spanned more than 160 languages including every official EU language, and the European deployment is operated end-to-end under European law, positioning the model for public-sector and regulated-industry buyers whose procurement requires exactly that provenance.

For builders, the practical Le Chonk read is straightforward. Coding agents gain a cheaper frontier-capable engine with three reasoning gears. Security operations gain the strongest defensible tool that can be owned outright. Regulated industries gain a frontier model with auditable lineage. And the API market gains a price floor pressure that every closed provider will feel.

For readers tracking how model capability is being metered by risk elsewhere, the gated-access contrast is instructive in our GPT-6 Cyber coverage, and the infrastructure angle of training frontier models on owned hardware connects to our Vera Rubin bandwidth analysis. The economics of token-efficient agents echo the trade-offs examined in our AI gateway comparison.

Deployment paths and what they cost

The release ships with four realistic deployment routes, and the right one depends on control requirements rather than capability preferences.

The Le Chonk API route through Mistral Studio or OpenRouter is the fastest start and the cheapest to try, at 70 cents per million input tokens and $2.10 per million output, with cached input at 7 cents. Cached input is the sleeper economics for agentic workloads, where system prompts, tool definitions, and repository context repeat across every request, and a 90 percent discount on the repeated portion compounds fast in long agent sessions.

The self-hosting route via Hugging Face weights served with vLLM or SGLang is the sovereignty play: full control, no per-token margins, and the model's built-in guardrails operating under the operator's own policies. The hardware bar is server-class, but organizations already running GPU capacity for other models can fold Le Chonk into existing fleets, and the FP8 and NVFP4 checkpoints are optimized for exactly that serving stack.

The managed-cloud route through Microsoft Foundry with private Azure deployment options splits the difference, offering enterprise platform features, regional deployment, and vendor accountability without operating hardware. The NVIDIA-hosted route via NIM container adds optimized serving for teams standardized on that stack, with NeMo AutoModel for customization and TensorRT LLM for throughput.

Customization cuts across all four. Mistral Forge tunes the model on proprietary data, policies, and detection logic, and Mistral's guidance notes that well-defined cybersecurity tasks can be distilled into smaller task-specific variants for higher accuracy at lower serving cost, which is the practical path for most security operations: run the big model for research, distill focused variants for the repetitive 90 percent.

Questions readers are asking

Can Le Chonk run on consumer hardware?

No, not meaningfully. A trillion-parameter mixture-of-experts model with roughly 50 billion active parameters needs multi-GPU server hardware even in FP8 or NVFP4 quantization. Self-hosters should plan for datacenter-class capacity, or use the API and hosted endpoints, which serve the same weights.

Is the license really open source?

It is open-weight rather than OSI-certified open source. Use, modification, fine-tuning, and commercialization are free; the single restriction targets businesses reselling the model as a hosted service, who need a separate commercial license from Mistral.

How does it compare with closed frontier models?

On coding agents it is competitive with the open-weight leaders and token-efficient enough to undercut closed APIs on running cost. On defensive cybersecurity it offers capability closed models withhold behind refusals. On general frontier reasoning, the closed leaders retain an edge in the hardest evaluations.

What should a security team do first?

Evaluate the model on your own vulnerability triage workload: reproduce known internal findings, test patch generation against your codebase, and measure refusal friction against your current tooling. The model's defensive strengths are exactly the tasks closed models most often block.

Sources
Subscription Form