Mistral Large 4 'Le Chonk': Open Weights, $0.70 API, One Catch

On October 29, Mistral kept the promise it made in early October: it published the weights of Mistral Large 4, the 1-trillion-parameter model it nicknames "Le Chonk." This is Mistral's largest open-weight release to date — and, on Mistral's own benchmark charts, it now leads the open-weight field in coding and defensive cybersecurity. But the parameter count is not the most interesting part of this story. The economics are: API pricing roughly halved overnight to $0.70 per million input tokens, the license is free for almost everything with one commercial carve-out, and the hardware needed to actually run the model still starts at a rack of datacenter GPUs. Here is what shipped, what it costs, and where the honest limits are.

Quick answer

• What: Mistral Large 4 ("Le Chonk") is a 1.05-trillion-parameter, natively multimodal mixture-of-experts model. As of October 29, 2026, its weights are downloadable from Hugging Face under Mistral's Ecosystem Open-Weight License.
• Cost: API access is priced at $0.70 per million input tokens, $0.07 for cached input, and $2.10 per million output — roughly half the preview rates. The weights are free, but self-hosting needs Blackwell-class GPUs.
• Catch: The license is a modified MIT license. You can use, fine-tune, and commercialize the model — unless you operate a model-as-a-service business, which needs a separate commercial license from Mistral.
• Status: The NVFP4 quantized checkpoint is live on Hugging Face (verified October 30); Mistral's own serving recipes also document an FP8 build. A full technical report is "coming soon."

What actually shipped

The headline specs, taken from Mistral's announcement and the model's Hugging Face card: 1.05 trillion total parameters with 52 billion active per forward pass — 49 billion per token, with the larger figure including embeddings and output layers. It is a granular mixture-of-experts design with 128 routed experts, four active per token, plus one shared expert (per the model card, via Unite.AI's read). A 1.6-billion-parameter Pixtral-style vision encoder handles variable-size images, and the context window is a full 1 million tokens — text and images in, text out.

Two engineering details are worth knowing. The model ships with an embedded DSpark speculative-decoding draft head that Mistral measured at up to 1.9x the throughput of the target model alone. And it now has three reasoning modes — Low, Medium, and High (the default) — so you can trade reasoning depth against token burn per request. It was trained from scratch on roughly 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own European datacenters, with native fluency in more than 160 languages.

What changed between the October 6 public preview and this release is three weeks of additional reinforcement learning. Mistral says those weeks drove "massive improvements in coding and defensive cybersecurity" — and, notably, that gains were still accelerating with no signs of saturation.

The benchmarks: read the chart, then the fine print

Mistral's announcement is heavy on numbers, and they are specific. On Cognition's private FrontierCode 2.0 benchmark — which tests whether a repository maintainer would actually merge an agent's change — Large 4 scores 19.7%, leading the open-weight field. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task. On Terminal-Bench 4.0 it reaches 41%, up 13 percentage points from the preview while cutting output tokens by 30%.

The human-preference signal is interesting too: in Surge Code's blind pairwise evaluations, graders preferred Mistral Large 4 over leading open-weight models including GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash — the subject of our own review of that model — with only Claude Opus 5.5 ranking ahead among the models tested.

On the defensive-security side: 51% on the Artificial Analysis Cyber Index (discover, reproduce, and patch vulnerabilities without breaking existing functionality), an average 96.6% refusal rate on malicious cyber requests across JailbreakBench, StrongREJECT, and AgentHarm — "the highest we measure among open models," Mistral says — and 97.1% attack resistance on the B3 Agent Security Benchmark, up from 93.3% in the preview. Enterprise metrics include 73.3% on FinWorkBench (second only to Kimi K3 among open models, in Mistral's words) and 84.5% on BrowseComp, up from a 75% starting checkpoint.

Here is the fine print, stated plainly because it matters: every benchmark above except one comes from Mistral's own announcement, and no independent lab has reproduced the vendor suite yet. The one independent data point is Artificial Analysis's own October 29 assessment: at its high reasoning setting, Large 4 scores 26 on the Intelligence Index v5, up from 25 for the preview — and AA measures $2.06 per Intelligence Index task, down from $5.54 for the preview, a 63% reduction (via VentureBeat). The same AA measurement puts DeepSeek V4.1 Flash 1017 at $0.60 per task, so the cheaper rival remains cheaper.

Pricing: the fine print

Mistral's official pricing, from its October 29 announcement: "Starting today, Mistral Large 4 is priced at $0.70 input / M, $0.07 cached input / M, and $2.10 output / M tokens." That is roughly half the preview's list prices ($1.36 / $0.14 / $4.18). One historical wrinkle: the preview also ran a 50%-off sale ($0.68 / $0.07 / $2.09) with no published end date, per the Developers Digest tracker — so this week's move looks more like making the sale price the standing price than a fresh discount. Artificial Analysis confirms the new rates in its assessment.

Two things that affect your bill. First, the reasoning mode: High is the default and burns the most tokens; Low replaces the old no-reasoning mode for simple tasks, Medium covers everyday work. Second, the route: Mistral's own API and Studio carry the headline rates, but third-party pass-through prices can differ — the LLM Reference tracker currently shows OpenRouter at $1.36 input / $4.18 output, the old list rate, so verify at checkout rather than assuming Mistral's number travels with the model. And if you self-host, the API price is moot but the GPU bill is not — see the hardware section below.

The license: open weight, not open source

The model ships under the Mistral Ecosystem Open-Weight License — a modified MIT license. Mistral's own wording: the model is "free to use, modify, fine-tune, and commercialize, with a single condition: businesses that operate a Model-as-a-Service offering must obtain a separate commercial license from Mistral before using the model commercially." VentureBeat's read of the license text adds that there is no revenue threshold — the trigger is whether you or an affiliate operate a model-as-a-service business, with some exceptions.

Two honest caveats. First, "open weight" is not "open source": you get the weights and the license, not the training code or data — so the "open science" recipe stays closed. Second, the announcement does not define where SaaS wrapping ends and model-as-a-service begins, as AI Weekly's coverage notes. If your product is anywhere near that boundary, get legal review before you build on the weights. The license is still dramatically more permissive than a closed API — but it is not the MIT you know.

What it takes to run it: the hardware math

This is the part that tempers the download excitement. The NVFP4 quantized checkpoint — the one we verified live on Hugging Face — targets Blackwell GPUs only (B200 / B300 / GB200 / GB300), which have dedicated hardware for the format. Per SGLang's documentation, the standard build runs on eight B300 GPUs, and a more heavily compressed variant runs on four GB300s (via goldenera.si's read of the docs). Other launch coverage cites "as few as four Nvidia B200 or B300 GPUs" for the quantized build (Crypto Briefing, ByteVyte) — treat the lower figure as the aggressive end of reported configurations, not a laptop spec.

The FP8 weights are roughly 1,050 GB by third-party arithmetic (systemdesign.academy) — a multi-GPU datacenter node, full stop. Mistral's own serving recipes ship for vLLM (docker pull vllm/vllm-openai:mistral-large-4, example config at tensor-parallel size 8), SGLang (example at tensor-parallel 4 on a Blackwell node), llama.cpp GGUF builds, and HF's remote-deployment endpoints. For most teams, the honest choice is the API or a managed endpoint — Microsoft Foundry is the first cloud provider offering private Azure deployment, with NVIDIA NIM containers and OpenRouter in the mix. Mistral also offers its Forge program for enterprises that want to fine-tune the model on proprietary data and policies.

The guardrails note: a model this capable is dual-use

Mistral is unusually candid in its announcement about the tension. It says it monitored cybersecurity-related use in live traffic during the preview: the model refused to comply with black-box attacks against live targets, but the company sees "a potential for uplift to conduct real attacks" and is modifying its guardrails accordingly, with more updates coming. That is the honest posture for an open-weight model marketed at defenders — the weights you can audit are the weights someone else can misuse.

The commercial expression of that posture: a specialized variant of Large 4, optimized for penetration testing, vulnerability assessment, and red-teaming, is going to enterprise design partners such as Orange Cyberdefense — with rigorous vetting, contractual safeguards, and testing restricted to owned or explicitly authorized systems in controlled environments. The public weights sit alongside it as the defensive-first distribution: strong on finding and patching flaws, refusing 96.6% of malicious cyber requests. Whether that balance holds in the wild is something only the coming weeks will show.

How it stacks up

On Mistral's own comparisons plus Artificial Analysis's independent cost measurement:

On cost per task — the only fully independent comparison — Artificial Analysis (via VentureBeat) measures $2.06 per Intelligence Index task for Mistral Large 4 at high reasoning, down from $5.54 for the preview (a 63% drop), versus $0.60 for DeepSeek V4.1 Flash 1017. If your priority is coding-agentic work inside an open-weight license, Mistral's case is token efficiency; if it is raw cost per task, the Chinese models still win.

For the broader landscape of how these coding agents are priced and limited in practice, see our roundup of the 7 best AI coding agents in 2026.

Pros and cons

Pros: Mistral's largest open-weight release to date; API pricing roughly halved with no published expiry; vendor-reported coding gains are specific and aggressive; 47% fewer tokens per task than its closest open-weight rival; defensive-cyber focus with refusal numbers stated openly; wide deployment ecosystem (vLLM, SGLang, llama.cpp, NVIDIA NIM, Microsoft Foundry); the sovereignty story is real for EU-regulated organizations.

Cons: every headline benchmark is vendor-reported and unreplicated; self-hosting is datacenter-class hardware, not a workstation; the license's model-as-a-service boundary is undefined enough to need legal review for borderline businesses; preview pricing history shows rates can move mid-cycle; the full technical report is still pending; guardrails on a dual-use model are visibly a work in progress.

What we couldn't verify

• Independent benchmarks. No independent lab has reproduced Mistral's coding or cyber benchmark suite yet — treat all vendor figures as claims, not confirmed performance.
• FP8 main-checkpoint status. We verified the NVFP4 quantized checkpoint live on Hugging Face on October 30; Mistral's serving recipes cover both FP8 and NVFP4. We did not independently confirm the main FP8 checkpoint's download state at press time.
• Third-party API prices. The OpenRouter figure ($1.36 / $4.18) comes from a tracker, not Mistral — check the live price at checkout before budgeting around it.
• The license boundary. Mistral's announcement does not define where SaaS wrapping ends and model-as-a-service begins. The license text is public — read Section 2 before you ship a product on the weights.

FAQ

Is Mistral Large 4 open source? No — it is open weight, which is not the same thing. You can download and run the model, but the training code, data, and recipe are not published. The license is a modified MIT with a model-as-a-service carve-out, not an OSI-approved open-source license.

Can I run it on my own hardware? Only on datacenter-class hardware: the NVFP4 build needs Blackwell GPUs (B200/B300/GB200/GB300), with documented configurations starting at 8x B300 for the standard build. For most teams, the API at $0.70 per million input tokens is the realistic way in.

How much cheaper is the API, really? The new rates — $0.70 input, $0.07 cached, $2.10 output per million tokens — are roughly half the preview's list prices. Artificial Analysis's independent measurement shows $2.06 per Intelligence Index task versus $5.54 for the preview, a 63% drop. It remains more expensive per task than DeepSeek V4.1 Flash 1017 at $0.60.

Is it better than GLM-5.3 or DeepSeek V4.1 Flash? On Mistral's own charts, it ties GLM-5.3 on the coding index while using 47% fewer tokens, and human graders in Surge Code's blind tests preferred it over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash. All of that is vendor-reported; independent reproduction is still pending.

What is the license catch, concretely? The model is free to use, modify, fine-tune, and commercialize — you only need a separate commercial license if you or an affiliate operate a model-as-a-service business. There is no revenue threshold; the boundary of what counts as MaaS is the part Mistral's announcement leaves undefined.

Key takeaways

• Mistral Large 4 ("Le Chonk") is downloadable from Hugging Face as of October 29, 2026 — Mistral's largest open-weight release to date.
• API pricing roughly halved: $0.70 input, $0.07 cached input, $2.10 output per million tokens, per Mistral's official announcement.
• The Ecosystem Open-Weight License is free for nearly everything but requires a separate commercial license for model-as-a-service businesses.
• Vendor-reported benchmarks lead the open-weight field in coding and defensive cybersecurity — independent reproduction is still pending.
• Self-hosting is datacenter-class (Blackwell GPUs, multi-GPU nodes) — for most teams, the API or a managed endpoint is the realistic way in.

Official links: Mistral's Large 4 open-weight announcement and the Hugging Face NVFP4 model card.

One trillion parameters, roughly half the API price, and a license you can read in full before you download — that is a genuinely good day for open weights. Now the independent benchmarks get the final word. October 30, 2026.