The download is the news
For three weeks, Large 4 was an open-weight model that nobody outside Mistral could weigh. The preview ran on Mistral's API, the Hugging Face page was a countdown, and the licence that would govern the weights had not been published. ALFW's October 6 piece ended on that gap.
The gap closed on October 29. The checkpoint is on Hugging Face as Mistral-Large-4-1T-A52B, with a quantized NVFP4 variant for Blackwell GPUs, and Mistral says it can be served with vLLM or SGLang. The model underneath is the one announced in October: 1.05 trillion parameters in total, 49 billion active per token, a 1M-token context, trained on 3,800 NVIDIA Grace Blackwell GPUs in Europe. What changed in the three weeks is additional reinforcement learning, three selectable reasoning modes, and a set of claims that can now be checked by people other than Mistral. One of those checks arrived the same day.
Two scoreboards, two winners
Mistral's launch charts place Large 4 first among open-weight models on Cognition's FrontierCode 2.0 at 19.7%, and first on Terminal-Bench 4.0 at 41%, up 13 points from the preview checkpoint. Read on its own, the release reads as a coding lead.
Artificial Analysis ran the released model independently for its Intelligence Index v5 assessment and got different numbers on the same named tests: 15% on FrontierCode 2.0, behind DeepSeek V4.1 Flash 1017 at 29% and GLM-5.3 at 22%, and 25% on Terminal-Bench 4.0. The difference is the harness and evaluation context around the model, which is precisely why a vendor chart cannot settle a ranking. On the composite Coding Agent Index, where Artificial Analysis does the comparing, Large 4 scores 48 and ranks second, one point behind GLM-5.3.
The overall picture is flatter still. Artificial Analysis gives the released model 26 on Intelligence Index v5, against 25 for the preview on the same scale. Seven open-weight models score higher; the leaders sit at 32. On Artificial Analysis's Intelligence Index v5, Large 4 scores one point above Muse Spark Open, making it the highest-scoring downloadable model developed outside China in that assessment. This is a narrow ranking within that index, not a claim that Large 4 leads every open-weight benchmark.
What genuinely improved
The measurable advance is economic. Artificial Analysis puts the cost of an Intelligence Index task at $2.06 for the released model, down from $5.54 for the preview. Average output tokens per task fell from 135,000 to 82,000. On the Coding Agent Index, Large 4 used 8.5 million total tokens per task against GLM-5.3's 15.9 million for a score one point lower. Mistral also halved its list prices on the same day, to $0.70 per million input tokens and $2.10 per million output tokens.
Cheaper rivals still exist within this assessment. Artificial Analysis measures DeepSeek's entry at $0.60 per Intelligence Index task while scoring it higher overall. That is a comparison under the index's evaluation setup, not a universal price for running either model in production. So the efficiency story is relative to Mistral's own preview, not to the field. A buyer choosing Large 4 is not choosing the cheapest capable open model, nor the highest-scoring one. The purchase is control: downloadable weights that can run on infrastructure the customer manages, a model trained in Mistral's European data centers according to its model card, and less dependence on Mistral's hosted API remaining available. That control comes with infrastructure, serving and maintenance costs of its own. Mistral's chief revenue officer made that pitch explicitly to VentureBeat this week, describing customers who want to own the model rather than rent continued permission to use it.
One more improvement deserves its own line. On Artificial Analysis's Omniscience test, the hallucination rate fell from 41% for the preview to 28%. Accuracy on the same test is 26%. A model that invents less while still answering only a quarter of factual questions correctly is easier to supervise, not yet reliable.
The licence, now readable
The October release published the term that matters most for anyone planning to build on the weights. The Mistral Ecosystem Open-Weight License is a modified MIT-style license, not the unmodified MIT license. According to VentureBeat's reading of the published terms, use, modification, fine-tuning and commercialization are permitted, but businesses operating qualifying model-as-a-service offerings must obtain a separate commercial license. The precise scope depends on the license's definitions and exclusions, which hosted-product developers should review before deployment. As VentureBeat's reading of the text sets out, the definition targets giving third parties inference or post-training access where those users control inputs, parameters or training data. It excludes internal use, and products where the model is embedded inside specific features rather than exposed as a service. There is no revenue threshold; the trigger is the kind of business, not its size.
That is a defensible line, and it is still a line. "Open weight" here means downloadable and broadly usable, with the one resale channel Mistral most cares about kept behind a negotiation. Teams planning a hosted offering on top of Large 4 should price that conversation into their plans before they build on the checkpoint.
The cyber claim carries its own footnote
Mistral positions Large 4 for defensive security work. Artificial Analysis gives it 51% on its Cyber Index, placing it third among open-weight models in the reported comparison. Separately, Mistral reports a 96.6% average refusal rate on malicious cyber requests, while its key benchmark figures include a 97.1% B3 Agent Security attack-resistance score. These are distinct evaluations and should not be read as equivalent measures of safety. The same announcement then states that monitoring of live traffic showed a potential for the model to assist real attacks, and that Mistral is modifying its guardrails in response.
Both halves belong in the record. The refusal figures describe the checkpoint as released and configured by Mistral; downloadable weights can be fine-tuned by whoever holds them, which changes what those figures guarantee in practice. Mistral keeps its stronger cyber variant behind vetting and contracts, available to partners such as Orange Cyberdefense for work on systems they own or are authorized to test. The public release and the controlled release are different products in effect, whatever the shared name. The existence of the vetted variant does not establish that the public checkpoint carries equivalent controls, particularly once users can modify the downloadable weights.
What to watch
Independent replication is now possible, which was not true three weeks ago. The tests to watch are replications of the FrontierCode and Terminal-Bench results under named harnesses, and the first hosted offerings either announcing a Mistral commercial licence or structuring themselves around the licence's embedded-product exclusion. If the ranking holds only inside Mistral's own charts, the coding-lead claim should be retired. If third parties reproduce it, Large 4 becomes the first non-Chinese open-weight model to hold that position with evidence behind it.
Sources
- [1] Mistral AI, "Mistral Large 4 (Le Chonk) is now open weight", October 29, 2026 (published 12:00 UTC, updated 15:42 UTC) — weights release, company's own benchmark figures, cyber monitoring admission, licence summary, API pricingRead source
- [2] Mistral AI / Hugging Face, model card mistralai/Mistral-Large-4-1T-A52B — 1T parameters, 49B active per token (52B including embeddings/output), 3,800 Grace Blackwell GPUs, NVFP4 checkpoint, DSpark draft headRead source
- [3] VentureBeat, "Mistral Large 4 'Le Chonk' arrives as the top open model outside China with improved coding performance and 63% lower cost-per-task", October 29, 2026 — Artificial Analysis October 29 assessment (Intelligence Index 26 vs preview 25, cost per task $5.54 to $2.06, FrontierCode 15% and Terminal-Bench 25% in AA's runs, Cyber Index components, Omniscience 28% hallucination / 26% accuracy), licence text reading (Section 2 MaaS condition, exclusions, no revenue threshold), CRO interviewRead source
- [4] Reuters via WTBX, "Mistral AI publicly releases Large 4, targets cybersecurity users", October 29, 2026 (14:03 UTC) — public release confirmation, Pierre Stock quotesRead source
- [5] ALFW, "Mistral's 'Le Chonk' Is a Trillion Parameters. You Can't Download It Yet.", October 6, 2026 — the preview piece this article follows upRead source