Mistral Large 4 on Hugging Face: Weights, License, vLLM Setup

Mistral released the Mistral Large 4 weights on October 29, 2026, and they are on Hugging Face now as mistralai/Mistral-Large-4-1T-A52B. The license is a modified MIT with one catch that matters if you resell inference: running a Model-as-a-Service business needs a separate commercial license from Mistral. Everyone else can use, modify, fine-tune and commercialize it. This is the follow-up to our preview-period Mistral Large 4 guide, which flagged the end-of-October weights date; the date held.

What the primary sources say#

Mistral's announcement says the weights are for an improved checkpoint, not the preview model. After about three weeks of extra reinforcement learning, Mistral reports (vendor-reported, all of it):

• Coding: 19.7% on Cognition's private FrontierCode 2.0, and 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3 while using 47% fewer tokens per task.
• Terminal work: 41% on Terminal-Bench 4.0, up 13 points from the preview and 30% fewer output tokens.
• Reasoning modes: low, medium and high (default), where low replaces the old "none" mode.
• Price: $0.70 input, $0.07 cached input and $2.10 output per million tokens on Mistral's own API.

The Hugging Face model card adds the details that decide whether you can run it: roughly 1T total parameters with 49B active per token (52B counting embeddings and output layers), 128 routed experts with 4 active, a 1M-token context window, text and image input with text output, and an embedded speculative-decoding draft head called DSpark that Mistral says gives up to 1.9x throughput.

One honest caveat on the headline claim. The model card says Mistral Large 4 scores 26 on the Artificial Analysis Intelligence Index v5, "the highest among western open-weight models", and states plainly that the leading Chinese open-weight models score higher on that aggregate. The pitch is coding, cyber and token efficiency, not overall intelligence.

The license, read carefully#

The license file is the Mistral Ecosystem Open-Weight License, which Mistral describes as a modified MIT license for commercial and non-commercial use. The only stated condition in the announcement and the model card: businesses that operate a Model-as-a-Service offering must get a separate commercial license before using it commercially.

What that means in practice:

This is not legal advice, and the license text is what binds you, so read it before you ship. The practical point is that third-party hosts of Mistral Large 4 will exist only if they strike a deal with Mistral. That is a different market from Kimi K3, where, as we noted in the preview guide, many providers compete on the same weights (see where to access Kimi K3 for that price spread). Expect fewer price wars on this model.

Run it: the verified command#

The model card recommends vLLM and gives this serving command, which we are reproducing from the card rather than from our own run. We have not tested it, because the model does not fit on anything we have.

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

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

Two details from the card that bite people. First, --tool-call-parser mistral and --reasoning-parser mistral are required for tool use and for separating reasoning from content; leave them off and your agent harness sees garbage. Second, add --speculative-config '{"method": "dspark", "num_speculative_tokens": 3}' to turn on the bundled draft head. SGLang has an equivalent recipe with --speculative-algorithm DSPARK, and a lmsysorg/sglang:mistral-large-4 image.

There is also an NVFP4 checkpoint (Mistral-Large-4-1T-A52B-NVFP4) that targets Blackwell GPUs on vLLM and SGLang, and GGUF builds via ggml-org and Unsloth for llama.cpp. Our own arithmetic, not Mistral's: 1T parameters at 16 bits is roughly 2 TB of weights, so even the 4-bit builds are a multi-GPU node, not a workstation. Use our VRAM guide to size anything smaller.

To point a coding agent at your server, Mistral's own CLI, Vibe, takes a local provider in ~/.vibe/config.toml with api_base = "http://<your-host-url>:8000/v1" and api_style = "reasoning". Any OpenAI-compatible harness works the same way, with the base URL swapped.

What a builder does Monday#

1. Do not self-host first. At $0.70/$2.10 per million tokens with $0.07 cached input, the API is the cheap way to find out whether the model suits your workload. A node that can hold a 1T-parameter model only pays off at sustained volume or when data residency forces your hand.
2. Add it to your router, not your stack. Run your own eval set against mistral-large-4 at medium reasoning effort and compare cost per finished task, because Mistral's pitch is fewer tokens per task. See model routers and the optionality advantage for why adding a model is cheap once you route.
3. Pick the use case it was built for. Mistral claims the model refuses 96.6% of malicious cyber requests while still doing defensive work like vulnerability triage. If you run security tooling in a regulated environment, that behavior profile is the reason to test it first.
4. Check the license against your business model before you build a hosted product on top of it.

What people are actually saying#

The Hacker News thread hit 1,360 points and 366 comments, and most of the oxygen went to one argument: is a model with a usage condition "open"? One commenter pointed to the license and argued "weights available" is the honest label, borrowing the old "source available" distinction. Another, who works at a university data center, said open weight should mean free to self-host, commercially, and for training. A third proposed a workaround: since the license blocks Model-as-a-Service, a GPU host could sell plain infrastructure and let clients bring the model. Whether that survives contact with Mistral's lawyers is untested.

The practical comments were more useful. One user who has used Mistral models almost exclusively since February said they measure worse "blank slate" than rivals, without skills, AGENTS.md or MCP, but hold up inside a configured harness. Several commenters made the sovereignty case: European firms want a contract with an EU company and EU-hosted KV cache, logging and billing, and a defense-contractor example turned on already having Mistral on an approved-vendor list.

The counter-case is the benchmark skepticism. One commenter warned that a model scoring well on benchmarks might be tuned for them, which is why every number above is labelled vendor-reported. Run your own tasks before you believe any of it.

FAQ#

Is Mistral Large 4 on Hugging Face?#

Yes. The weights are at mistralai/Mistral-Large-4-1T-A52B, with an NVFP4 variant for Blackwell GPUs. Mistral says you can also use it through Mistral Studio, the API (mistral-large-4) and OpenRouter.

Is Mistral Large 4 really open source?#

It is open-weight under the Mistral Ecosystem Open-Weight License, a modified MIT. It is free to use, modify, fine-tune and commercialize, except that Model-as-a-Service businesses need a separate commercial license. Whether that qualifies as open source is exactly what the HN thread argued about.

How much does Mistral Large 4 cost on the API?#

$0.70 per million input tokens, $0.07 cached input and $2.10 output, according to Mistral's announcement. The preview ran at a discounted $0.68/$2.09 during a launch sale.

Can I run Mistral Large 4 locally?#

Only with a multi-GPU node. The model card's example uses tensor parallel size 8 on vLLM. GGUF builds exist for llama.cpp, but a 1T-parameter model is not a laptop workload.

Continue Reading#

• Mistral Large 4: Specs, Pricing, Benchmarks and Weights Date - the preview-period guide that predicted this release
• Kimi K3 Open Weights on Hugging Face - the other big open release and how its hosting market played out
• Where to Access Kimi K3: Every Provider and Price Compared - what a crowded open-weight host market looks like
• Model Routers and the Optionality Advantage - why adding a model to a router is cheap
• How Much VRAM Do You Need for a Local LLM in 2026 - size hardware before you download anything
• Best Open-Source Models - Mistral and the rest of the open-weight field, ranked
• Best Local Models - what you can self-host, by memory tier

Sources#