Le Chonk is loose: 1 trillion parameters, your PC is crying

In short

• The weights are on Hugging Face, three weeks after the October 6 preview: about 1,000 billion parameters, of which 49 to 52 billion work on each token.
• According to Mistral: 48% on the AA Coding Agent Index, 51% on the AA Cyber Index. Artificial Analysis showed neither as of 10/29 at 14:39 UTC.
• API at $0.70 per million input tokens, $2.10 output. "Modified MIT" license: a company that sells access to models (itself or its affiliates) must sign separately with Mistral before any commercial use.

What came out

After the October 6 preview and "three weeks of additional reinforcement learning", here's the card:

• Size: a Mixture-of-Experts of about 1,000 billion parameters; 49 billion active per token, 52 billion counting the input and output layers (the repo name says "A52B");
• Experts: 4 out of 128 on each token, plus 1 shared expert;
• Context: 1 million tokens; text and image inputs; three reasoning modes.

Trained "from scratch" on 3,800 Nvidia Grace Blackwell GPUs, in Mistral's data centers in Europe. Mistral plays the European "sovereign AI" card to the hilt, with nineteen testimonials (Cisco, Nvidia, Orange Cyberdefense…).

The in-house numbers

All measured or chosen by Mistral, final version:

• FrontierCode 2.0 (Cognition): 19.7%, top spot;
• AA Coding Agent Index: 48%, "effectively tied" with GLM-5.3, with 47% fewer tokens per task;
• Terminal-Bench 4.0: 41%, 13 points more than the preview (so 28%, our math);
• AA Cyber Index: 51%; refusal of malicious cyber requests: 96.6%.

In Surge's blind tests quoted by Mistral, it's preferred to GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, behind Claude Opus 5.5.

The Hugging Face card admits the rest: on Artificial Analysis's general index (v5), 26, "the highest among western open-weight models", but the best Chinese open models do better. I like it when a lab writes that itself.

What independent measurements say

Artificial Analysis site read on 10/29/2026 at 14:22 then 14:39 UTC.

• Final version: already in their menu, but its page returns a 404 error. No readable score, not the 48%, not the 51%, not the 26 Mistral attributes to them.
• The preview, on the other hand, is measured: 25 on the general index v5 (median of comparable models: 42), and very talkative: 150 million tokens for the whole index, versus 64 million median.

We're watching.

What actually changes

The price is cut in half. The preview was listed at $1.36 per million input tokens and $4.18 output. Since 10/29: $0.70 and $2.10 ($0.07 cached), prices confirmed by the API docs at 14:39 UTC. If you send 10 million tokens and get 2 million back: $11.20 instead of $21.96.

The weights. The repo weighs about 1,061 GB, in FP8. The example commands spread it over 8 GPUs ("a node of 8 GPUs"), without saying which: about 133 GB of weights per card, before working memory. An NVFP4 version of the weights is also published on Hugging Face, lighter, for Blackwell GPUs. Your laptop can go to bed.

The license, without overselling it

It's the "Mistral Ecosystem Open-Weight License", "a modified version of the MIT license": you can use, copy, modify, redistribute, sell, retrain. The condition that changes everything:

• if you or one of your affiliated companies runs a "Model as a Service" business, you need a separate commercial license with Mistral before any commercial use;
• "Model as a Service": giving a third party access to the model (API, hosted endpoint) while letting it control the inputs, settings or training data;
• not covered: a product where the model only serves specific features, simply relaying to a model hosted elsewhere, internal use (nothing made available to third parties) and use through Mistral's own products.

In short: not plain MIT.

What we don't know yet

• The guardrails: Mistral saw in its traffic a "potential for uplift to conduct real attacks" and is changing its protections "in the months to come". Nothing says whether this will affect the weights already published.

Key terms

• Open weights: the model files can be downloaded. That says nothing about the license.
• Mixture of experts (MoE): a model split into sub-networks, of which only a few work on each token.
• Active parameters: the share that computes for each token. It affects speed, not memory: the 1,000 billion still have to fit somewhere.

About the source

• Who announces it: Mistral AI, post "Mistral Large 4 (Le Chonk) is now open weight".
• Date: published on 10/29/2026 at 12:00 UTC, modified at 14:18 UTC; Hugging Face repo modified at 14:11 UTC.
• Links: Mistral post · Hugging Face card · license · model docs · NVFP4 version · Artificial Analysis, preview

Our math

• R1: sum of the sizes of the 67 files in the repo (Hugging Face API, 10/29 14:21 UTC): 1,060,613,093,988 bytes, i.e. about 1,061 GB. The weights index gives 1,060,548,427,328 bytes.
• R2: 1,061 GB ÷ 8 GPUs ≈ 132.6 GB, rounded to 133 GB per GPU, weights only.
• R3: 10M × $0.70 + 2M × $2.10 = $7.00 + $4.20 = $11.20; at the preview's listed price: 10M × $1.36 + 2M × $4.18 = $13.60 + $8.36 = $21.96.
• R4: 41% − 13 points = 28% for the preview on Terminal-Bench 4.0.

Source passages

Show the 20 sources

Ref.	Quote (original language)	URL
S1↑1↑2↑3↑4	"Earlier this month, we released Mistral Large 4 (Le Chonk), our flagship 1-trillion-parameter, natively multimodal model, in public preview." / "Today, we're releasing the weights of this improved Mistral Large 4."	https://mistral.ai/news/mistral-large-4-oss/
S2↑1↑2↑3↑4	"a 1 trillion-parameter natively multimodal Mixture-of-Experts model with 49 billion active parameters per token (52 billion including embeddings and output layers), trained from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's own datacenters in Europe."	https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B
S3↑1↑2	"~1T total / ~52B active parameters: granular Mixture-of-Experts with 128 routed experts (4 active per token) and 1 shared expert." / "1M context length" / "Multimodal input: accepts text and image input, with text output."	https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B
S4↑1↑2↑3↑4↑5	"This license is a modified version of the MIT license." / "(a) use, copy, modify, merge, publish, distribute, sublicense, and/or sell copies of the Model; (b) run, deploy, further train or otherwise modify the Model" / "If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose." / "\"Model as a Service\" means giving a third party access to the Model or its derivative works for inference or post-training by any means (for instance via API or hosted endpoints), in a manner that allows such third party to exercise meaningful control over the inputs, parameters, or training data. This does not include (a) end-user products with model capabilities solely embedded within specific features or harnesses, or (b) mere relaying of requests to models hosted by others."	https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE
S5↑	"The requirements set forth in Section 2 do not apply to: (a) internal use of the Model or its derivative works, defined as any use that does not make the Model or its derivative works available to third parties; or (b) any use of the Model accessed through Mistral's official products."	https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/blob/main/LICENSE
S7↑1↑2↑3	"Starting today, Mistral Large 4 is priced at $0.70 input / M, $0.07 cached input / M, and $2.10 output / M tokens."	https://mistral.ai/news/mistral-large-4-oss/
S8↑1↑2↑3↑4↑5↑6+1	"Mistral Large 4 ranks as the best open-weight model for coding" / "Just three weeks of additional reinforcement learning" / "On Cognition's private FrontierCode 2.0 benchmark [...] Mistral Large 4 leads with a score of 19.7%." / "On the Artificial Analysis Coding Agent Index, Mistral Large 4 scores 48%, effectively tied with the leading open-weight model, GLM-5.3, while using 47% fewer tokens per task on average." / "On Terminal-Bench 4.0, Mistral Large 4 improved by 13 percentage points to 41% compared to its public preview, taking the lead among open-weight models while reducing output tokens by 30%."	https://mistral.ai/news/mistral-large-4-oss/
S9↑1↑2↑3	"Mistral Large 4 scores 51% on the Artificial Analysis Cyber Index" / "achieving an average refusal rate of 96.6% on malicious cyber requests, the highest we measure among open models."	https://mistral.ai/news/mistral-large-4-oss/
S10↑	"On the Artificial Analysis Intelligence Index v5, Mistral Large 4 scores 26, the highest among western open-weight models. The leading open-weight models from China score higher on this aggregate index"	https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B
S12↑	"Mistral Large 4 now supports three reasoning modes" / "Low" / "Medium" / "High (default)"	https://mistral.ai/news/mistral-large-4-oss/
S14↑	"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. We are modifying our guardrails accordingly, and will continue updating them in the months to come."	https://mistral.ai/news/mistral-large-4-oss/
S16↑1↑2	"--tensor-parallel-size 8" / "--tp-size 8: example tensor-parallel size for a node of 8 GPUs; adjust to your topology." / "A quantized NVFP4 checkpoint [...] targets Blackwell GPUs"	https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B
S17↑1↑2	"datePublished":"2026-10-29T12:00:07.000Z"; "dateModified":"2026-10-29T14:18:44.772Z" (post); "lastModified":"2026-10-29T14:11:16.000Z" (Hugging Face repo)	https://mistral.ai/news/mistral-large-4-oss/
S18↑1↑2	"October 6, 2026" "Public Preview" / "It features 52B active parameters and 1.05T total parameters, and a 1.6B vision encoder." / "$1.36 Input/M Tokens" "$0.14 Cached input/M Tokens" "$4.18 Output/M Tokens"	https://docs.mistral.ai/models/mistral-large-4-0
S19↑1↑2↑3	"Mistral Large 4 Preview scores 25 on the Artificial Analysis Intelligence Index, placing it at the lower end among comparable models (median: 42). When evaluating the Intelligence Index, it generated 150M tokens, which is very verbose in comparison to the median of 64M." / "Pricing for Mistral Large 4 Preview is $1.36 per 1M input tokens [...] and $4.18 per 1M output tokens"	https://artificialanalysis.ai/models/mistral-large-4-preview
S20↑	Reading of 10/29/2026 14:22 UTC: https://artificialanalysis.ai/models/mistral-large-4 returns HTTP 404; the site menu lists {"slug":"mistral-large-4","name":"Mistral Large 4 (High)"} with no score; neither "48%" (Coding Agent Index) nor "51%" (Cyber Index) on readable pages.	https://artificialanalysis.ai/models/mistral-large-4
S21↑1↑2	Index des poids : "total_size": 1060548427328; 54 fichiers "consolidated-000NN-of-00054.safetensors"; config : "quant_method": "fp8"	https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B/tree/main
S22↑	"In Surge Code's blind pairwise evaluations, Mistral Large 4 was preferred over leading open-weight models including GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash, ranking behind only Claude Opus 5.5 among the models evaluated."	https://mistral.ai/news/mistral-large-4-oss/
S24↑1↑2↑3↑4	Reading of 10/29/2026 14:39 UTC: docs.mistral.ai shows "$0.7" "Input" / "$0.07" "Cached input" / "$2.1" "Output" (header still "October 6, 2026" "Public Preview"); https://artificialanalysis.ai/models/mistral-large-4 still returns HTTP 404, neither "48%" nor "51%" on the page.	https://docs.mistral.ai/models/mistral-large-4-0
S25↑	"id":"mistralai/Mistral-Large-4-1T-A52B-NVFP4", "private":false, "gated":false (API Hugging Face, 29/10/2026 14:39 UTC) / README : "A quantized NVFP4 checkpoint [...] is also available, with a smaller memory footprint, faster inference and same performance. It targets Blackwell GPUs"	https://huggingface.co/mistralai/Mistral-Large-4-1T-A52B-NVFP4