Mistral Large 4 open weights: how to download and run it

As of this morning you can download Mistral Large 4 and run it yourself: the full weights are on Hugging Face at mistralai/Mistral-Large-4-1T-A52B, and the license allows commercial use. It's also enormous. The official FP8 checkpoint is built for an 8-GPU server, and even the smallest community build, a 1-bit quant from Unsloth, needs about 256 GB of memory.

When Mistral put out the API preview on October 6, the weights were only a promise for the end of the month (our first look at Mistral Large 4 covers that launch and the pricing confusion around it). Mistral hit that date.

What Mistral posted

The announcement came from @MistralAI at 14:21 UTC: "Today, we release the weights of a new and improved Le Chonk." This build isn't the preview. Mistral calls it an improved version tuned for agentic coding and cybersecurity, and says it ranks first among open models on Cognition's FrontierCode benchmark. The post also says security vendors including CrowdStrike, Cloudflare, Rubrik and Zscaler tested it for cyber defense.

The tooling followed within the hour. @UnslothAI posted quantized GGUF files and claimed the model "runs on a Mac Studio with our 1-bit GGUF." Unsloth co-founder @danielhanchen gave the figure I'd want before downloading anything: the 1-bit version keeps 91% of top-1 accuracy on coding, agent tasks and math while coming in about 80% smaller than the 1.1 TB FP8 original. Then @vllm_project announced day-0 support, with FP8 and NVFP4 checkpoints and a built-in speculative decoding drafter.

@testingcatalog repeated Mistral's headline claim that Large 4 "ranks as the best open-weight model for coding," then said they hope a new Mistral Small comes soon. Plenty of small businesses would get more use out of that one.

The numbers, from Mistral's model card

Every figure below is Mistral's own, from the release post and the Hugging Face model card. I haven't seen independent checks of them yet.

If you own a business, read the license closely. Commercial use, modification and fine-tuning are all allowed. If you sell the model itself as a hosted service, you need a separate commercial license from Mistral. Running it inside your own company or product is fine.

How to run Mistral Large 4

Which path makes sense depends on the hardware you have.

1. Use the API. If you only want to try it, Mistral's API at the prices above is the cheapest way to test, and you don't download anything.

2. A GPU server with vLLM. The model card's examples assume one 8-GPU node. The official vLLM container runs it like this:

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

You get an OpenAI-compatible endpoint on port 8000. Mistral recommends temperature=1.0 and top_p=0.95, and you can set reasoning_effort to low, medium or high (the default).

3. A big-memory workstation with llama.cpp. Unsloth's guide lists every quant. The 1-bit UD-IQ1_S is 233 GB and should fit in 256 GB of unified memory. The 2-bit UD-Q2_K_XL is 378 GB, and 4-bit is 605 GB. Its server command:

./llama.cpp/build/bin/llama-server \
  -hf unsloth/Mistral-Large-4-1T-A52B-GGUF:UD-Q2_K_XL \
  --ctx-size 32768 --temp 1.0 --top-p 0.95 \
  --chat-template-kwargs '{"reasoning_effort":"high"}' \
  --host 127.0.0.1 --port 8080

Swap the tag to UD-IQ1_S for the 1-bit build. Unsloth doesn't publish speed figures, so I wouldn't expect chat-speed replies from a 1-bit trillion-parameter model on a desktop.

My read

This matters most to firms that can't send data to someone else's cloud, like a law office or a company bound by European data rules. For a typical small business, the math hasn't changed. Hardware that holds even the 1-bit build costs thousands of dollars, the API runs $0.70 per million input tokens, and the closed models are still ahead on general office work.

What I find more interesting is how fast open models now arrive with working tooling. Unsloth and vLLM had support ready on day one, and NVIDIA shipped a NIM container for it, the same pattern we saw with Nemotron 3.5 Transcribe this week. "Can we run this ourselves?" now has a real answer, even if for most firms it's "yes, but you don't need to."

If you're trying to work out where a model like this fits in your own workflows, private or hosted, New Face Design's free process audit maps that out before you buy any hardware.