What is Mistral Large 4?
Mistral Large 4 is Mistral AI's flagship open-weight model, a mixture-of-experts with about 1 trillion total and 52 billion active parameters, text and image input and a 1M-token context. Use it for coding agents or defensive security work you want to host yourself or run cheaply through the API. Model-as-a-Service businesses need a separate commercial licence.
What can you do with it?
Mistral aims the model at agentic coding first. It reports 48% on the Artificial Analysis Coding Agent Index, level with GLM-5.3 while using 47% fewer tokens per task on average, and 41% on Terminal-Bench 4.0. A coding agent built on it takes a repository and an issue and returns a change for review; Mistral's own Vibe agent runs on it.
The second use is defensive security. Mistral reports 51% on the Artificial Analysis Cyber Index, which covers finding, reproducing and patching vulnerabilities, and a 96.6% average refusal rate on malicious cyber prompts. On the Intelligence Index v5 the model card gives it 26, which Mistral calls the top western open-weight score while noting that leading Chinese open models score higher.
Try it on a small task
Suggested workflow: before downloading a trillion-parameter checkpoint, call it through the Mistral API as mistral-large-4. Give it one real issue from your repository with the relevant files and ask for a diagnosis and a minimal patch, once at medium and once at high reasoning effort.
Inspect whether the patch passes your tests, whether the explanation matches the actual cause, and how many output tokens each effort level used. If the results hold up and your data must stay in-house, self-hosting the weights is the next step.
Pricing and access
Mistral lists API prices of $0.70 per million input tokens, $0.07 for cached input and $2.10 for output. The model is also offered in Mistral Studio, the Vibe coding agent, OpenRouter and as an NVIDIA NIM container, and enterprises can fine-tune it with Mistral Forge.
The weights sit on Hugging Face as mistralai/Mistral-Large-4-1T-A52B under the Mistral Ecosystem Open-Weight License, a modified MIT licence. Use, modification, fine-tuning and commercial use are allowed, but a company running a Model-as-a-Service business must sign a separate commercial licence first. The model card's vLLM example spreads the model over eight GPUs, so self-hosting usually means a multi-GPU server; according to Hugging Face's launch post, the 2-bit GGUF quantizations run on a single Apple Silicon Ultra computer with 512 GB of unified memory.
What stands out
- About 1T total and 52B active parameters, 128 routed experts with 4 active
- 1M-token context; text and image input, text output
- Reasoning effort low, medium or high (default)
- FP8 and NVFP4 checkpoints on Hugging Face
Before you choose
- Model-as-a-Service providers need a separate commercial licence.
- Full-precision self-hosting needs a multi-GPU server; only 2-bit GGUF builds fit one 512 GB Apple Silicon Ultra machine.
- Leading Chinese open models score higher on the overall Intelligence Index, per Mistral.
Go to the source
Sources & further reading
Based on the official documentation below. Features and access can change; check the provider before making a decision.
- Mistral AI: Mistral Large 4 is now open weight (opens in a new tab)
- Mistral Docs: Mistral Large 4 (opens in a new tab)
- Hugging Face: Mistral-Large-4-1T-A52B model card (opens in a new tab)
- Mistral Ecosystem Open-Weight License (opens in a new tab)
- Hugging Face blog: Welcome Le Chonk (opens in a new tab)
Was this page useful?
Help us make our coverage more useful. This is feedback on the page, not a rating of the tool.
Go a little further