How to try Mistral Large 4: API, free demos or vLLM

How to try Mistral Large 4: call it through Mistral's API as mistral-large-4, or download the open weights from Hugging Face and serve them with vLLM on a multi-GPU server. This guide follows the model card as read on October 30, 2026. We did not run the commands.

Mistral Large 4 is a 1 trillion-parameter multimodal mixture-of-experts model, which uses only part of its network for each token: 49 billion active parameters, or 52 billion counting embeddings and output layers. It reads text and images and writes text, with a 1 million token context. We cover the release in Mistral releases Mistral Large 4 weights.

Option 1: use the Mistral API

The model card says the model is served on the Mistral API as mistral-large-4. This is the easiest route, because a model this size needs a large GPU server. The card itself recommends the API "if local serving is subpar."

Option 2: try the free demos

The card links a demo gallery, Le Chonk Gallery, with Spaces built on the model. It says official demos run on the Mistral API with "a small shared free budget per visitor for the first week of the release." Mistral published the open weights on October 29, so that free budget is short-lived. You can also bring your own Mistral API key.

Option 3: serve it yourself with vLLM

Mistral recommends vLLM and ships a Docker image. Pull it first:

docker pull vllm/vllm-openai:mistral-large-4

Then start the server. The card's example uses eight GPUs:

docker run --runtime nvidia --gpus all \
  -v ~/.cache/huggingface:/root/.cache/huggingface \
  -p 8000:8000 \
  --ipc=host \
  vllm/vllm-openai:mistral-large-4 \
  --model mistralai/Mistral-Large-4-1T-A52B \
  --tensor-parallel-size 8 \
  --tool-call-parser mistral \
  --reasoning-parser mistral \
  --enable-auto-tool-choice \
  --gpu-memory-utilization 0.85

The two parsers matter. The card says --tool-call-parser mistral and --reasoning-parser mistral are required for tool use and to separate reasoning from the answer. If you run out of GPU memory, lower --max-model-len or --max-num-seqs.

The card also lists an NVFP4 checkpoint, a smaller version that targets Blackwell GPUs and works with vLLM and SGLang. Other routes are SGLang, Transformers and llama.cpp, with GGUF files published by the ggml-org project.

Send your first request

With the server running on port 8000, point the OpenAI SDK at http://localhost:8000/v1 with the API key set to EMPTY, then call client.chat.completions.create. Set reasoning_effort to low, medium or high. High is the default. Mistral recommends temperature=1.0 and top_p=0.95 at every effort level, and notes that max_completion_tokens counts reasoning tokens as well as the visible answer.

Which option fits you

Pick the API if you want to test quality on your own prompts today without hardware. Pick the demos for a quick look at coding, vision, tool calls and JSON output, and expect the free budget to end soon. Pick vLLM if data has to stay on your own machines or you plan to fine-tune. The card lists Axolotl, Unsloth and TRL as fine-tuning routes, and Unsloth publishes its own guide and GGUF files.

What the model offers once it runs

The card lists three reasoning efforts, low, medium and high, so one model can handle a quick chat reply or a long agent run. It supports function calling and strict JSON schema output, where the response must match a schema you give. It also handles images, charts and documents. Mistral says it is especially strong on cybersecurity, finance and law, and that it scores 26 on the Artificial Analysis Intelligence Index v5, the highest among western open-weight models, while Chinese open-weight models score higher on that index.

Check the license first

The weights are not plain open source. The card calls the license a modified MIT license for commercial and non-commercial use, but running a Model-as-a-Service business needs a separate commercial license. Our license explainer has the detail.

Common problems

• Out-of-memory errors: lower context length and concurrency, as above.
• Reasoning text mixed into answers, or tools not firing: the parser flags are missing.
• Slow answers: try reasoning_effort="low" for chat-style replies.