Mistral Large 4 Scores 26 on the Model Card and 38.4 in Preview, With a Tighter Licence Than Large 3
A.I. / news
Mistral Large 4 Scores 26 on the Model Card and 38.4 in Preview, With a Tighter Licence Than Large 3
The 1-trillion-parameter open weights are on Hugging Face, but the index score, context window and licence all differ from what October's preview coverage said.
Mistral's Large 4 weights are on Hugging Face under a licence that bars selling the model as a service without a separate Mistral deal, and the model card's headline score is 26 on Artificial Analysis's Intelligence Index v5. An October preview write-up of the same model gave 38.4. The two figures sit on different pages and neither explains the gap.
The model card lists about 1 trillion total parameters, 49 billion active per token, or 52 billion counting embeddings and output layers. It describes 128 routed experts with four active per token, one shared expert, and a Pixtral-style vision encoder. Input is text and images, output is text. The card gives no release date.
What changed between the preview and the card
Trending Topics reported Artificial Analysis's preview results on Oct. 6, when the weights were still unreleased and the evaluator listed the model as proprietary. The card now carries different numbers and different terms.
| Item | Oct. 6 preview write-up | Model card now |
|---|---|---|
| Intelligence Index | 38.4 | 26 (v5) |
| Context window | about 524,000 tokens | 1M tokens |
| Licence | none named | Mistral Ecosystem Open-Weight License |
| Ranking claim | 8th among open-weight models | highest among western open-weight models |
Both pages agree on one thing: leading Chinese open-weight models score higher. The card says so directly. The preview article listed seven of them above Large 4, led by Xiaomi's MiMo-V2.6-Pro at 46.3 and Z.ai's GLM-5.3 at 44.8.
The article as fetched gives no index version number, so the two scores cannot be placed on one scale. Treat 38.4 and 26 as measurements of different things until Artificial Analysis publishes both on one page.
- MiMo-V2.6-Pro46.3 points
- GLM-5.344.8 points
- Kimi K343.6 points
- Mistral Large 4 (preview)38.4 points
- Motif 333.6 points
Source: Trending Topics, reporting Artificial Analysis, accessed 2026-11-06
The licence is the real change
The card calls the licence a modified MIT and says it covers commercial and non-commercial use. Operating a Model-as-a-Service business needs a separate commercial licence from Mistral. The card metadata lists the licence as "other."
Trending Topics reported that Mistral's Large 3 shipped under Apache 2.0 and that Medium 3.5 carried a modified MIT excluding companies above $20 million in monthly revenue. Large 4 is therefore less permissive than its predecessor for anyone reselling inference. For a company using the weights inside its own product, the card's text imposes no cap.
Compare Linum's Apache 2.0 text-to-image release, which carries no resale clause.
Cost per task and the verbosity problem
Artificial Analysis found Large 4 used about 200 million output tokens to run its full index, against a median of 81 million for comparable models. It called the model very verbose. The full run cost about $1,600. Cost per index task was $1.13, against $2.01 for GLM-5.3 and $2.00 for Kimi K3.
API pricing in the preview was $1.36 per million input tokens and $4.18 per million output tokens. The card's own claim runs the other way: on Artificial Analysis's Coding Agent Index it reports a 48% score, effectively tied with GLM-5.3, with 47% fewer tokens per task. That is Mistral's reading of a coding index, not the same measurement as the index full run.
Closed frontier models sat further out in the preview. Claude Opus 5.5 scored 57.6, OpenAI's GPT-6 Astra 52.7 and Google's Gemini 4 Argon 52.6.
What it takes to run
The card gives no VRAM figure or minimum GPU count. Its serving examples use eight GPUs: vLLM with tensor parallelism of 8, SGLang with a tensor-parallel size of 8, and a torchrun launch across eight processes. A separate NVFP4 checkpoint targets Blackwell GPUs. Mistral says it trained the model from scratch on 3,800 NVIDIA Grace Blackwell GPUs in its European datacentres.
One more inconsistency sits on the card. Its text claims fluency in 160-plus languages, while the page metadata shows 39.
The card also lists GGUF quantisations for llama.cpp, Ollama and LM Studio. It recommends temperature 1.0 and top_p 0.95, and says to lower the maximum model length or sequence count if memory runs short.
Sources
More in A.I.
- 01John Langford Says He Left AI Frontiers After Microsoft AI ReorgMicrosoft's lab page confirms the Sept. 14 move into Microsoft AI, while Langford's blog post says the new owner's philosophy is scaling and not research.
- 02Nemotron 3.5 Transcribe Posts 4.62 Word Error Rate Across Seven Languages, NVIDIA's Own NumberThe offline speech model has 4.3 billion active parameters, 40 languages and a licence that permits commercial use.
- 03Cloudflare's Security Audit Skill Has an Open Issue Saying a Made-Up Finding Can Pass ValidationThe MIT-licensed skill runs six phases of agent audits. Its tracker shows 43 open issues, including a profile-wipe report and a Windows failure.
- 04Anthropic's Agent Memory Study Puts Opus 5.5 at 0.838 With Memory and 0.655 WithoutThe Nov. 6 post tests notes, wikis and offline dreaming, and reports where a wiki beat search and where it did not.