Three weeks after putting Mistral Large 4 into public preview, Mistral AI released the full weights of the 1 trillion parameter model on October 29, and the announcement page reads less like a model card and more like a pricing argument. The company's headline claim on coding is not that Le Chonk scores higher than GLM-5.3 on the Artificial Analysis Coding Agent Index. It scores 48%, effectively tied. The claim is that it reaches that score while using 47 percent fewer tokens per task on average. In a market where every lab now quotes benchmark percentages that land within a few points of each other, token counts are becoming the axis where models actually separate, and this release is Mistral planting a flag there.
The Release in Context
Mistral announced the open-weight release on its own blog, framing it as a sovereign AI play: trained from scratch on European infrastructure, downloadable from Hugging Face, deployable on your own hardware under a modified MIT license. SiliconANGLE reports the training ran on 3,800 Nvidia Grace Blackwell GPUs in Mistral's European data centers, and the model carries 1 trillion total parameters with roughly 52 billion active per token, which matches the repository name mistralai/Mistral-Large-4-1T-A52B on Hugging Face.
A few details from the announcement matter for how you should read the numbers below. The released weights are not the preview weights. Mistral ran three additional weeks of reinforcement learning after the October 6 preview, and the company attributes most of the coding and cybersecurity gains to that stretch. The model ships with three reasoning modes (low, medium, high), the same compute dial pattern we saw with Exa's compute dial and Gemini 4's tiered reasoning: one model, three operating points on the cost-performance curve. And the license, a modified MIT called the Mistral Ecosystem Open-Weight license, is free to use, modify, fine-tune, and commercialize, with one carve-out: businesses operating a Model-as-a-Service offering need a separate commercial agreement.
What the Numbers Actually Say
Almost every headline number here is vendor-reported, on benchmarks Mistral selected. We checked what independent signals exist and label them in the table. The exception is structural: we pulled the model's config.json straight from the Hugging Face repo, and those architecture facts are as verified as anything in this story.
| Metric | Value | Source status |
|---|---|---|
| Cognition FrontierCode 2.0 | 19.7%, first place among listed models | Vendor chart |
| AA Coding Agent Index | 48%, tied with GLM-5.3 | Vendor chart, index is third-party |
| Tokens per task vs leader | 47% fewer on average | Vendor claim |
| Terminal-Bench 4.0 | 41%, up from 28% in preview, 30% fewer output tokens | Vendor chart |
| Surge blind pairwise, code | Wins vs GLM-5.3 and Kimi K3 (64%), loses to Claude Opus 5.5 (22%) | Third-party eval, vendor-charted |
| AA Cyber Index | 51% | Vendor chart |
| Malicious cyber refusal rate | 96.6% average across three evals | Vendor claim |
| BrowseComp | 84.5%, from a 75% SFT checkpoint | Vendor claim |
| B3 Agent Security | 97.1% attack resistance | Vendor claim |
| AA general intelligence index | Fifth among open models | Independent, via Hugging Face reporting |
Two rows deserve emphasis. First, the Surge Code results cut both ways: graders preferred Le Chonk's code over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, but ranked only Claude Opus 5.5 above it, with a 22 percent win rate in that pairing. So the honest summary is top open model outside China on code, still behind the closed frontier. Second, on Artificial Analysis's overall intelligence index, the reporting around the release places it fifth among open models, not first. Mistral leads the categories it trained hard on (coding and defensive security) and trails elsewhere. That is exactly what a post-preview reinforcement learning sprint produces: a sharpened spike, not a lifted floor.
Why This Matters: Cost Per Task Is the Benchmark Now
Here is the data scientist's read. When two models tie on a score of 48 and one needs half the tokens to produce the same accepted output, the benchmark score has stopped being the interesting variable. Token efficiency is a direct multiplier on cost, latency, and context-window exhaustion, and in agent loops those effects compound across dozens of tool calls per task.
Run the arithmetic with Mistral's own API pricing ($0.70 per million input tokens, $2.10 per million output, cached input at $0.07). Assume an agentic coding task where the weaker-token-efficiency model emits 200,000 output tokens and Le Chonk emits 106,000 for the same accepted result, which is the claimed 47 percent reduction applied to output only, our assumption, not Mistral's stated example. At $2.10 per million output tokens, that is $0.42 versus $0.22 per task: a saving of $0.20 on a single task, $200 per thousand tasks, before you touch the input side. Multiply by the fact that Mistral separately claims a 30 percent output-token reduction on Terminal-Bench between preview and release, and the company is compounding the same lever twice in three weeks: capability up, tokens down, in the same update.
The architecture explains why this is structurally achievable rather than a fluke. The config.json on the Hugging Face repo shows a sparse mixture-of-experts design: 128 routed experts plus 1 shared expert per MoE layer, 4 experts active per token, across 72 layers, with a 6,144 hidden size and MLA-style split attention heads (64 query heads, separate rope and nope dimensions, a yarn rope type). Roughly 5 percent of the parameters fire per token. Sparse activation is what lets a 1 trillion parameter model reason at the marginal compute cost of a ~52B dense model, and reinforcement learning on top of that can trim tokens by teaching the model to stop rambling, not just to answer correctly. This is the same open-weights MoE economics we traced in Meta's Muse Spark Open release and in DeepSeek V4.1 Flash's cost analysis: frontier-class scores increasingly come from routing, not density.
There is also a quieter engineering tell. The config's quantization_config reports fp8 as the shipping quantization, and the repository's safetensors files total about 1,060 GB. That is a trillion parameters in 8-bit floating point: Mistral is no longer treating fp8 as an optimization users perform, but as the default delivery format for frontier open weights. Compare that with the Kimi K3 pricing story, where the whole argument was cost per task falling faster than price per token. Le Chonk's release pushes the identical thesis from the European side.
Running It Is Still a Hardware Problem
Sovereignty marketing does not change memory bandwidth. Here is the self-hosting math with every assumption labeled. The fp8 release files total roughly 1.06 TB, so native-fp8 serving needs at least that much memory in fast tiers, which means multi-GPU nodes or a server-class unified-memory box. The llama.cpp and Unsloth community 2-bit builds compress the download to about 233 GB, and per Hugging Face's own testing reported by Martin Cid, the only Apple machine that fits a 2-bit build is a 512 GB Mac Studio, a $22,299 configuration. And speed is the real constraint: Hugging Face measured about 15 tokens per second on an older M3 Ultra at 2 bits, fine for conversation, painful for long agent runs, with output dropping to roughly 3 tokens per second on a four-RTX-6000-Pro rig under very long contexts. Squeezing to 2 bits also costs accuracy relative to the fp8 weights Mistral benchmarked. If you were weighing this against API pricing, our local versus API cost breakdown still gives the simple rule: self-hosting a trillion-parameter MoE is a compliance and sovereignty purchase, not a cost reduction. The Mac Studio M5 Ultra hardware math covers what that box can and cannot hold.
The Dual-Use Fine Print
Mistral's cybersecurity pitch is unusual: a frontier open-weights model tuned so it answers defensive work (reproduce a vulnerability, validate it, patch it) without refusal, while hitting a claimed 96.6 percent average refusal on malicious cyber requests. Hugging Face additionally reports 82 percent on a task requiring the model to reproduce and fix a real open-source vulnerability, and says several closed frontier models score near zero there because they refuse the whole task. Booz Allen, Cisco, CrowdStrike, Cloudflare, and ENISA all contributed quotes to the launch.
Then the announcement says something most labs would bury: while monitoring live preview traffic, Mistral observed "a potential for uplift to conduct real attacks" and is tightening guardrails accordingly. That caveat has a hard limit. Once weights are public, guardrails are a courtesy to the careful, not a wall; anyone, including attackers, can download the checkpoint and retune it. Sophos' Rafe Pilling called it "clearly a dual-use technology" in the press coverage, and he is right. The defensive value is real and verifiable (auditable weights, on-prem deployment, no vendor terms-of-service risk), and so is the offensive one. This is the same tradeoff the ecosystem keeps losing the argument about in our benchmark skepticism guide: openness distributes capability in both directions at once.
Outlook
Three things to watch. First, the promised technical report: the RL recipe behind +13 Terminal-Bench points in three weeks, with fewer tokens, is the part every lab will want, and its absence keeps the gains unfalsifiable for now. Second, independent evaluation on the coding index: ties reported by the vendor's own charts are the first thing outside labs will re-measure, and this piece in the open-model race already has the shape of last week's Gemini 4 carbon result, a lead measured in heads, not chasms. Third, the MaaS license clause: whether cloud providers sign commercial deals or route around them decides if "open weights" here means open distribution or open distribution with a toll booth.
For builders, the practical takeaway is narrower than the press release. If your workload is agentic coding where you pay per token and evaluate accepted diffs rather than chat quality, Le Chonk is the first open Western model whose headline spec is cost per completed task, and the first one you can actually audit. Run it against your own repos on OpenRouter before believing any of the charts. That, incidentally, is what Mistral itself asks for at the end of the announcement, and it is the one line in the post we would not edit.