llama.cpp llm hugging-face open-weights mistral mistral-large-4 le-chonk moe

Mistral Large 4 Open Weights Ship: 1-Trillion-Parameter "Le Chonk" Lands on Hugging Face

Posted by Reda Fornera on 2026-10-31
Estimated Reading Time 16 Minutes
Words 2.6k In Total

The Eiffel Tower and Paris skyline — stock imagery representing the French origins of Mistral AI and its European-built trillion-parameter open-weights release, not an actual screenshot of the Hugging Face repository

Countdown over: Mistral Large 4 open weights are live

After weeks of watching a Hugging Face repo that everyone knew existed, the Mistral Large 4 open weights are finally downloadable. Mistral released the open weights of Mistral Large 4 — the trillion-parameter model the company has been calling “Le Chonk” — as mistralai/Mistral-Large-4-1T-A52B on Hugging Face, with repo metadata showing the big push landing on October 29, 2026. It closes the loop on the promise made when the model entered public preview earlier in the month: weights at the end of the month, right on schedule.

What did you actually get? The main repository is the FP8 checkpoint: 54 consolidated safetensors shards of roughly 20 GB each, plus config, tokenizer, and processor files — about 67 top-level entries and, by our tally of the shard sizes, roughly 1.06 TB of downloads. Alongside it sits an official NVFP4 checkpoint (mistralai/Mistral-Large-4-1T-A52B-NVFP4) targeting NVIDIA Blackwell GPUs with a smaller memory footprint and “same performance,” per the model card. Community GGUF conversions came fast too: from Unsloth at unsloth/Mistral-Large-4-1T-A52B-GGUF, and — under ggml-org — the llama.cpp team’s own repo. (If you stumble across a repo called Mistral-Large-3-675B-Instruct-2512-NVFP4 while browsing, note that it’s from the previous Large generation — created back in late 2025 — not part of this release. For that earlier generation in detail, see our Mistral 3 deep dive.)

Server racks in a datacenter — stock imagery representing the terabyte-scale FP8, NVFP4 and GGUF checkpoints described above, not an actual file listing from Hugging Face

The license is the Mistral Ecosystem Open-Weight License, which the LICENSE file describes as “a modified version of the MIT license.” The gist: you can use, copy, modify, merge, publish, distribute, sublicense, sell copies, and fine-tune — commercial and non-commercial — with one carve-out. If you or your affiliates operate a Model as a Service business, you need a separate commercial license from Mistral before commercial use. The definition is narrower than you might fear: giving third parties inference access in a way that lets them “exercise meaningful control over the inputs, parameters, or training data” is where the MaaS line sits, while end-user products with the model embedded in specific features, mere relaying to models hosted by others, and internal use all sit outside Section 2. For almost everyone actually reading this blog, the practical answer is: download, deploy, fine-tune, sell it as part of your product — fine. Resell tokens from it as a hosted endpoint — talk to Mistral first.

One regulatory question the release leaves open, which CSO Online’s coverage raises: under the EU AI Act, models trained with more than 10²⁵ FLOPs are presumed to pose systemic risk, and nothing in Mistral’s announcement says whether the Mistral Large 4 open weights cross that threshold or what risk assessment preceded the release. Pierre Stock told CSO that Mistral is a full signatory of the EU GPAI Code of Practice, and ENISA said in a statement published alongside the announcement that it is testing the model — but the systemic-risk question itself remains unanswered by the release materials.

What’s under the hood of the Mistral Large 4 open weights

Le Chonk is a natively multimodal Mixture-of-Experts model. The model card headline: approximately 1 trillion total parameters with 49 billion active per token (52 billion including embeddings and output layers), and a 1M-token context window. It takes text and images as input and generates text, using a Pixtral-style vision encoder — 1.6B parameters, 30 layers — trained to handle variable image sizes and aspect ratios. It’s fluent in 160+ languages, including every official language of the EU. And it was trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral’s own datacenters in Europe, which is the line doing a lot of work in the sovereignty pitch.

The HF launch blog goes deeper on the architecture, and it’s worth a skim because this isn’t a vanilla MoE:

  • Text decoder: 72 layers, hidden dimension 6,144, 64 attention heads.
  • Two kinds of sparsity at once. The MoE router picks 4 of 128 routed experts per token, plus 1 shared expert that always runs — 5 active experts per token. Separately, attention uses Multi-head Latent Attention (MLA) combined with a DeepSeek Sparse Attention (DSA) indexer that selects up to 2,048 key positions for each query. One sparsity selects the feed-forward networks, the other selects context positions.
  • Attention geometry quirk: a 3:1 RoPE-to-NoPE pattern — three decoder layers use RoPE, every fourth layer uses NoPE — with KV state shared across layers, and “LeRope,” a YaRN-scaled RoPE with a learned multiplier per frequency.
  • Layer mix: the first decoder layer uses a dense FFN; the other 71 are MoE blocks.
  • Day-0 speculative decoding: the checkpoint ships with an embedded DSpark draft head — 3 dense MLA draft layers in BF16 proposing blocks of 3 tokens — which Mistral measured at up to 1.9x higher throughput without changing outputs.

What the Mistral Large 4 open weights improved: three extra weeks of RL

This is the story that will matter most to anyone who tried the preview and filed it away as “impressive, but watch it mature.” Between the public preview and the weights drop, Mistral kept scaling reinforcement learning, and it shows up most clearly in terminal-based coding.

On Terminal-Bench 4.0, the announcement reports the model improved by 13 percentage points — from 28% in the preview (charted as self-reported) to 41% — which the post says takes the lead among open-weight models, while also cutting output tokens by 30%. On Cognition’s private FrontierCode 2.0 benchmark, which scores end-to-end code quality on whether a repo maintainer would actually merge the change, Mistral Large 4 claims the lead at 19.7%. 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. (One vendor-internal wrinkle: the HF launch blog reports a combined 49.8% on the same index, ahead of DeepSeek V4 Pro 0813 and Qwen3.8 Max, while the announcement and model card report 48% — the two tier-A sources disagree by 1.8 points depending on reasoning configuration. Quote either number with that caveat.) And in Surge Code’s blind human preference evaluations, graders preferred Mistral Large 4’s code over GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, with the model ranking behind only Claude Opus 5.5 among those evaluated.

One honesty check before you copy those numbers into a slide deck: third-party snapshots don’t fully agree with vendor-reported scores. The HF launch post, citing Artificial Analysis figures as of October 29, lists Mistral Large 4 at 28.3% on Terminal-Bench 4 and the fifth-best open-weights model on the AA Intelligence Index; the HF model card adds the caveat that “the leading open-weight models from China score higher on this aggregate index.” Independent harnesses, different quantizations, different dates — the gap may be entirely mundane. But the gap itself is the reminder: “leads FrontierCode and Terminal-Bench” is a Mistral-reported claim until independent re-runs settle it. The full technical report is still pending.

The pricing, at least, is plain: the Mistral API serves mistral-large-4 at $0.70 per million input tokens, $0.07 per million cached input tokens, and $2.10 per million output tokens — aggressive pricing for a flagship, and clearly intended as the convenient alternative to babysitting a terabyte of weights.

The security elephant in the room

Here’s the part of the story that makes the Mistral Large 4 open weights release genuinely unusual. Mistral’s own announcement includes a cyberattack-uplift warning — in the same post that ships the download links.

“Though we observed the model refusing to comply with black-box attacks against live targets, we see a potential for uplift to conduct real attacks.”

The company says it is modifying the model’s guardrails accordingly and will keep updating them “in the months to come.” And there’s the rub, as CSO Online’s coverage makes explicit: those updates won’t reach the copies organizations pull from Hugging Face. The article quotes Pierre Stock, Mistral’s vice president of science: “Subsequent safeguard updates are applied to our managed API and future model releases, rather than automatically updating downloaded checkpoints.” Downloaded weights are frozen-in-time weights. That’s a feature for reproducibility and a fact of life for open distribution — but it means the safety story diverges the moment the download counter starts ticking.

The other half of the tension is that Mistral is pitching the model for security work. On the Artificial Analysis Cyber Index, the announcement reports 51%, and the HF blog highlights the reproduce-then-patch test — reproduce a real vulnerability in open-source software, then patch it — where Mistral Large 4 achieves 82%, the highest score the blog reports on that test within the index. The defense argument is practical: fixing vulnerabilities requires reproducing them, and reproduction looks like an attack to a closed-source filter. Mistral is also gating the most offensive-capable usage behind a vetted channel: the announcement says vetted design partners such as Orange Cyberdefense are getting preview access to a specialized variant of Mistral Large 4 optimized for penetration testing, vulnerability assessment, and red-teaming — “subject to rigorous vetting and contractual safeguards,” with testing restricted to owned or authorized systems. The HF blog also notes that several closed frontier models score near zero on the same test because they refuse to try, and that the model solves 93% of 40 Cybench challenges.

Mistral is upfront that the index crown isn’t total: its own chart shows DeepSeek-V4.1-Flash and MiMo-V2.6-Pro scoring higher on the Cyber Index, a caveat CSO Online’s own reporting echoes. On the refusal side, the announcement claims an average refusal rate of 96.6% on malicious cyber requests (the cyber-related prompts of JailbreakBench, StrongREJECT and AgentHarm) — the highest Mistral measures among open models — plus 97.1% attack resistance on the B3 agent-security benchmark, up from 93.3% at preview.

The analyst reactions in CSO’s piece are the counterweights worth keeping in mind. A Gartner principal analyst, Niyati Daftary, reads the 96.6% narrowly: it tells you what the model does in Mistral’s test scenario, “not how it will behave with your data, your adversarial use cases or in production in your environment against motivated attackers.” Forrester’s Charlie Dai points out that without a published technical report, “best open-weight model” is hard to validate across workloads. Both are right, and both are the standard tax on any frontier-capability claim made via blog post.

What the release means for the open-weights landscape

Strip away the benchmark tables and this is a European infrastructure story. A ~1T-parameter frontier model, trained on 3,800 of NVIDIA’s latest-generation GPUs in the company’s own European datacenters, with what Mistral describes as “clear lineage from inception to release,” released under a permissive-with-one-carve-out license on the same day the API went live. La Banque Postale’s statement in the announcement makes the buy-side pitch explicit — sovereignty “is not optional” for a public bank — and Microsoft announced that Mistral Large 4 is available through Microsoft Foundry, with NVIDIA also offering hosted endpoints and a NIM container.

The competitive shape is also interesting: this is a Western open-weight release that concedes the aggregate intelligence crown to Chinese open models while claiming leadership in coding and cybersecurity specifically. (The head-to-head numbers against the current open-weight leaders deserve their own article — consider this the setup, not the verdict.) What’s not arguable is the pattern: the open-weights frontier keeps moving, and the gap between “something you can self-host” and “the best thing you can rent” keeps shrinking at the same time companies like Mistral try to have it both ways — open distribution for sovereignty, managed API for revenue.

For enterprise deployers, the CSO takeaway stands: the Mistral Large 4 open weights shift responsibility for patching, monitoring, and misuse prevention to you. Run it as a controlled component, not an autonomous agent.

How to download and run Mistral Large 4 open weights

If you have a B200/B300-class box or a node of Blackwell GPUs, the recommended path is vLLM serving the NVFP4 checkpoint:

1
2
3
4
5
6
7
8
vllm serve mistralai/Mistral-Large-4-1T-A52B-NVFP4 \
  --config-format mistral \
  --load-format mistral \
  --tensor-parallel-size 8 \
  --enable-expert-parallel \
  --dtype bfloat16 \
  --max-model-len 131072 \
  --gpu-memory-utilization 0.85

The FP8 checkpoint of the Mistral Large 4 open weights is served via the dedicated vllm/vllm-openai:mistral-large-4 Docker image — though note the FP8 model card’s serve command isn’t identical to the NVFP4 flags above: it adds --tool-call-parser mistral, --reasoning-parser mistral, and --enable-auto-tool-choice, and omits --config-format/--load-format, so check the card rather than copying flags across variants. SGLang has a day-0 recipe, and transformers supports it via AutoModelForMultimodalLM (with AutoProcessor handling image inputs) across 8 GPUs via torchrun. Reasoning effort is configurable per request — low, medium, or high (default high) — and note that reasoning can’t be fully disabled. One caveat from the model cards applies to everything local: “For optimal performance, we recommend using the Mistral AI API if local serving is subpar.” Even on the described hardware, the vendor’s own suggestion is that renting may beat self-hosting on quality.

A computer monitor displaying lines of colorful source code — stock imagery illustrating a workstation running local commands like the llama.cpp GGUF serve command below, not an actual terminal screenshot of the Mistral Large 4 quantization

The genuinely fun part is llama.cpp. The llama.cpp team’s GGUF repo spans IQ2_XXS, Q2_K, Q4_K_M, NVFP4, Q8_0, and BF16. The 2-bit quants reportedly run on a single Apple Silicon Ultra box with 512 GB of unified memory:

1
llama serve -hf ggml-org/Mistral-Large-4-1T-A52B-GGUF:Q2_K

The HF blog’s reported ballpark for that config on an M3 Ultra: about 90 tok/s prefill and 15 tok/s generation. Not fast, but “a trillion parameters on one workstation” is a sentence that will age well. (If you want more background on running big open models on local hardware, see our guide to running Gemma 4 locally and our zero-cost local agentic AI setup on a DGX Spark.) If you’d rather not own the problem at all, the model is on the Mistral API, OpenRouter, Microsoft Foundry, and Inference Providers — and day-0 fine-tuning support arrived in TRL, Axolotl, and Unsloth.

Grab the Mistral Large 4 open weights, run your own evals, and — as Mistral itself invites — share the results. With a terabyte-scale release this fresh, independent numbers are the ones that will matter.

References and further reading

Please let us know if you enjoyed this blog post. Share it with others to spread the knowledge! If you believe any images in this post infringe your copyright, please contact us promptly so we can remove them.

FEATURED TAGS
llama.cpp llm hugging-face open-weights mistral mistral-large-4 le-chonk moe
// adding consent banner
ESC