Mistral Large 4 open weights: 1T-parameter model, licence, benchmarks, how to run it

Mistral AI released the weights of Mistral Large 4, nicknamed "Le Chonk", on 29 October 2026: a roughly 1-trillion-parameter multimodal mixture-of-experts model with a 1M-token context window that anyone can download from Hugging Face under a modified MIT licence. It is the biggest open-weight model a Western lab has shipped, and the company says it is tied with the Chinese open model GLM-5.3 on coding agents while using far fewer tokens. Running it is another matter: the practical entry point is a 512 GB Apple Silicon machine at 2-bit precision, or a small rack of Nvidia Blackwell GPUs.

Key points

• Mistral Large 4 is a multimodal mixture-of-experts model of about 1 trillion parameters (49B to 52B active per token, depending on the page), released as open weights on 29 October 2026 after a hosted preview that began on 6 October.
• The licence is the Mistral Ecosystem Open-Weight License, a modified MIT licence; anyone offering the model as a service needs a separate commercial licence from Mistral.
• Mistral reports 41% on Terminal-Bench 4.0 (28% at preview) and 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task on average. These are company-reported figures.
• Hugging Face lists a 1M-token context window and quantised builds; a Q2_K GGUF runs on one Apple Silicon Ultra with 512 GB of memory at about 15 tokens per second of generation.
• Hosted API pricing on Mistral's page is $0.70 per million input tokens and $2.10 per million output tokens.

What happened

Mistral published the weights for Mistral Large 4 on Hugging Face on 29 October, three weeks after it opened a hosted public preview on 6 October and said the downloadable weights would follow "by the end of October" after more safety work. Hugging Face published its own integration guide the same day, and the New Stack and The Next Web both reported the release.

The model is natively multimodal (text and images in, text out) and, per Mistral, fluent in more than 160 languages. It has three reasoning-effort settings, low, medium and high, with high as the default. Hugging Face's guide adds that reasoning cannot be switched off.

Mistral says it trained the model from scratch on 3,800 Nvidia Grace Blackwell GPUs in European data centres. The weights are published as an FP8 checkpoint (mistralai/Mistral-Large-4-1T-A52B) and an NVFP4 checkpoint for Blackwell GPUs, with GGUF files from ggml-org and Unsloth for llama.cpp.

Mistral's launch post leads with coding and security: "Built for Coders, Trusted by Defenders." It reports a 51% score on the Artificial Analysis Cyber Index and a 96.6% average refusal rate on malicious cyber requests, which it calls the highest among open models measured. It also reports 97.1% attack resistance on the B3 agent security benchmark, up from 93.3% at preview.

Background

Mistral, the Paris company founded in 2023, built its reputation on small open models and moved to larger releases under a permissive licence with Mistral Large 3. Large 4 changes the terms: the Apache 2.0 licence used earlier gives way to a modified MIT licence with a clause for model-as-a-service businesses. For most developers, companies running it internally and researchers, nothing changes; for cloud providers who want to resell it, a deal with Mistral is required.

The release lands in a crowded field. Mistral compares itself with GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, all Chinese open-weight models. Pierre Stock, Mistral's VP of Science, told the New Stack's reporter that no company should depend on a provider that can cut off access, and argued that open weights cannot be withdrawn once published. CEO Arthur Mensch said other labs have raised hundreds of billions of dollars for comparable capabilities while Mistral raised roughly $6 billion in three years; The Next Web notes that claim is hard to verify.

The numbers

Everything below is reported by Mistral unless noted, so treat it as a vendor claim until independent evaluations catch up. On Terminal-Bench 4.0, an agentic coding test, The Next Web lists the comparison like this:

Mistral Large 441

GLM-5.340

DeepSeek V4.1 Flash31

Mistral Large 4 preview28

Kimi K321

Terminal-Bench 4.0 score (agentic coding) (%)

	%
Mistral Large 4	41
GLM-5.3	40
DeepSeek V4.1 Flash	31
Mistral Large 4 preview	28
Kimi K3	21

Source: The Next Web, citing Mistral's reported comparison

The 13-point rise over the preview came with 30% fewer output tokens, Mistral says. On the Artificial Analysis Coding Agent Index it scores 48%, effectively tied with GLM-5.3, and uses 47% fewer tokens per task on average. That matters for cost: Stock argues per-token prices mislead when a model needs fewer tokens for the same job.

Other results Mistral cites: 73.3% on FinWorkBench, second only to Kimi K3 among open models, and 84.5% on BrowseComp. In blind pairwise tests on Surge Code it says users preferred the model to GLM-5.3, Kimi K3 and DeepSeek V4.1 Flash, behind only Claude Opus 5.5. The Hugging Face model card reports an Artificial Analysis Intelligence Index v5 score of 26, the highest among Western open-weight models, though it says several Chinese open models score higher on that aggregate.

Sources disagree on a few details. Mistral's page says 1 trillion parameters; Hugging Face's summary says 49B active per token and 52B including embedding and output layers, while The Next Web says 52B. The Hugging Face guide lists Terminal-Bench at 28.3% and B3 at 93.3%, which are the preview numbers; Mistral's launch post gives the updated 41% and 97.1%.

How it works

The text decoder has 72 layers. The first is dense; the other 71 use a mixture of experts that routes each token to 4 of 128 experts, plus one shared expert. Attention uses Multi-head Latent Attention with a DeepSeek Sparse Attention indexer that picks up to 2,048 key positions per query, and KV state is shared across layers. A 1.6B-parameter Pixtral vision encoder handles images.

Running it. Mixture-of-experts models need memory for all parameters even though only about 5% are active per token, so the weights are the constraint, not the compute. Per Hugging Face:

• llama.cpp on a Mac: IQ2_XXS and Q2_K fit on one Apple Silicon Ultra with 512 GB unified memory. Q2_K on an M3 Ultra gives about 90 tokens per second prefill and 15 generation. Larger quants need distributed inference.
• Workstation GPUs: four RTX 6000 Pro cards with NVFP4 at 32K context manage roughly 10 tokens per second of generation.
• Servers: the vLLM recommended setup uses eight GPUs with tensor parallelism and the NVFP4 checkpoint on Blackwell, with an optional speculative decoding head. Hugging Face's Inference Endpoints recipe uses eight H200s with FP8 weights, supporting up to 131K context and 16 concurrent requests.
• Fine-tuning: Hugging Face shows LoRA training of the FP8 model on eight B300 GPUs with TRL.

For the general steps of pulling and quantising an open model, see our guide to running an open-weight LLM locally.

What it means

For developers, this is the first trillion-parameter open model from outside China that is competitive on coding agents, with the weights, quantised builds and inference recipes available on day one. Teams that cannot send code to a third-party API, such as banks, defence suppliers and European public bodies, gain a model they can host themselves.

For Mistral's rivals, the token-efficiency claim is the one to test. If 47% fewer tokens per task holds up, the effective cost gap with GLM-5.3 is larger than the near-tie on accuracy suggests. For cloud providers, the licence clause means they cannot simply resell it.

Independent signals are still thin. Mistral's own comparisons dominate this launch, and Hugging Face's model card places the model fifth among open-weight models on the Artificial Analysis aggregate index, so "best open-weight model for coding" is Mistral's wording, not a settled ranking.

What we don't know yet

• Whether the benchmark gains hold up in independent testing; most figures come from Mistral, and the Terminal-Bench 4.0 comparison is Mistral's table as reported by The Next Web.
• The exact active-parameter count: 49B or 52B depending on the page and how embeddings are counted.
• The full text of the licence's acceptable-use terms; the model card states only a general ban on infringing third-party rights.
• When the smaller distilled versions, including one meant for laptops, will arrive; The Next Web says they are expected "in coming months" with no committed dates.
• Whether an API price of $0.70 and $2.10 per million tokens stays; The Next Web said pricing had not been published, while Mistral's page now lists it.

What to watch next

• Independent runs on Artificial Analysis, Vals and Terminal-Bench that confirm or revise Mistral's numbers.
• Quantisation work that makes the model usable on less than 512 GB.
• The smaller distilled models Mistral has promised.
• How cloud providers respond to the model-as-a-service licence clause.

Sources

• Mistral AI: Mistral Large 4 (Le Chonk) is now open weight
• Hugging Face blog: Welcome Le Chonk, Mistral Large 4
• Hugging Face model card: Mistral-Large-4-1T-A52B
• The Next Web: Mistral releases open weights for its 1 trillion-parameter Large 4 model
• The New Stack: Mistral's Large 4 open weights arrive with a 1-point coding lead over GLM-5.3
• Let's Data Science: Mistral opens public preview of Large 4