Mistral releases Large 4 weights with 1 trillion parameters
Mistral released improved Large 4 weights on October 29, following the model's earlier public preview. Developers can customize and host the multimodal model, provided their hardware and intended use fit its deployment and licensing requirements.
Spec sheet
- Total parameters
- 1 trillion
- Active per token
- 49 billion; 52 billion including embeddings and output layers
- NVFP4 weights
- Approximately 610 GB, before runtime memory
- FP8 weights
- Approximately 1.06 TB, before runtime memory
- License
- Mistral Ecosystem Open-Weight License
- Release
- 29 October 2026
The download is only the start
Large 4 uses a mixture-of-experts architecture. Only part of the model participates in each token calculation, which separates its active parameter count from its full weight-storage requirement. The release gives organizations control over a model that handles images and text. That control comes with decisions about quantization, serving software and the accelerator capacity assigned to each workload.
Size the serving configuration
The vLLM recipe lists four GB300 GPUs for NVFP4 and sixteen H200 GPUs for FP8 among its validated configurations. It validates context lengths up to 256K tokens, below the model card's advertised 1M maximum. Longer inputs and more simultaneous requests need their own memory budget. A deployment plan should include runtime state alongside the stored weights, then measure throughput with the actual input lengths customers send.
Compare the work, then the license
Mistral positions Large 4 against GLM-5.3 on agentic coding. Token consumption deserves attention because long coding runs can generate substantial intermediate work. Buyers should compare completed tasks, failure recovery and serving costs on the same workload. The license is another decision point: Large 4 carries separate commercial rules for Model-as-a-Service offerings. DeepSeek's V4.1 Flash 1017 card lists MIT licensing. Those terms matter when a team plans to sell model access, even if both models offer downloadable weights.
Claim check
- Confirmed The model card specifies 1T total parameters and 49B active per token, or 52B including embeddings and output layers. [3]
- Company claim Mistral reports 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens. [3]
- Unclear Independent quality and total serving cost on a shared production workload are not established by the vendor comparison. [3]
Sources
- Mistral Large 4 (Le Chonk) is now open weight · Mistral AI · 29 Oct
- Welcome Le Chonk: Mistral Large 4 on Hugging Face, October 29 · Hugging Face · 29 Oct
- Mistral Large 4 model card · Mistral AI · 31 Oct
- Mistral Large 4 deployment recipe · vLLM · 31 Oct
- GLM-5.3 model card and evaluation methods · Z.ai · 31 Oct
- DeepSeek V4.1 Flash 1017 model card and license · DeepSeek · 17 Oct