Mistral Large 4 Is Open Weight. Read the License and Do the Memory Math Before You Self-Host

2026-11-03来源 dev.to OpenSourceEnglish

On October 29, 2026, Mistral released the weights of Mistral Large 4 (nicknamed "Le Chonk") on Hugging Face. It's a roughly 1 trillion parameter, natively multimodal Mixture-of-Experts model with a 1M-token context, and Mistral calls it the best open-weight model for coding.

"Open weight" headlines make self-hosting sound like a download away. Two details decide whether it actually is for your team: what the license lets you do, and how much hardware "52B active" really needs. This post walks through both, plus how I'd decide between the API and running it yourself.

Benchmark figures below are Mistral's own reported numbers unless noted. Treat them as claims until independent evals land.

What was released

  • Model: mistralai/Mistral-Large-4-1T-A52B on Hugging Face, in two official variants: FP8 and NVFP4. Unsloth and the llama.cpp team published GGUF quants.
  • Architecture: ~1T total parameters, granular MoE with 128 routed experts (4 active per token) plus 1 shared expert. 49B parameters are active per token (52B if you count embeddings and output layers). Artificial Analysis lists 1.05T total.
  • Inputs: text and images in, text out, with a Pixtral-style vision encoder. 1M-token context. Function calling, and reasoning effort configurable per request (Low, Medium, and High as the default).
  • Training: from scratch on 3,800 NVIDIA Grace Blackwell GPUs in Mistral's datacenters in Europe.
  • Serving: day-0 support in transformers, vLLM and llama.cpp; Mistral also points to SGLang.
  • API price: $0.70 per 1M input tokens, $0.07 cached input, $2.10 output.

Mistral's reported results include 19.7% on Cognition's private FrontierCode 2.0, 41% on Terminal-Bench 4.0 (up 13 points from the preview while using 30% fewer output tokens), 48% on the Artificial Analysis Coding Agent Index (roughly tied with GLM-5.3 while using 47% fewer tokens per task), 84.5% on BrowseComp, and 97.1% attack resistance on the B3 Agent Security benchmark.

Detail 1: "Modified MIT" is not MIT

The license is the Mistral Ecosystem Open-Weight License, a modified MIT license. It grants broad rights to use, copy, modify, fine-tune, distribute and sell the model and derivatives. Then section 2 adds the condition that matters:

If the Licensee or any of its affiliates operates a Model as a Service business, the Licensee must enter into a separate commercial license with Mistral before using the Model or its derivative works for any commercial purpose.

"Model as a Service" means giving a third party access to the model for inference or post-training (an API, a hosted endpoint) in a way that lets them exercise meaningful control over the inputs, parameters, or training data.

The license carves out:

  • Internal use: any use that doesn't make the model available to third parties.
  • End-user products where the model is "solely embedded within specific features or harnesses".
  • Relaying requests to models hosted by others.
  • Use through Mistral's own products.

In practice, a quick way to read it:

What you're building Likely bucket
Internal coding assistant or RAG over company docs Internal use, no extra license
SaaS feature: "summarize this ticket" button Embedded in a specific feature
Public API where customers send their own prompts and params Model as a Service, needs a commercial license
Fine-tuning service for customers' data Model as a Service (post-training)

Note the condition is about whether you or an affiliate operate a MaaS business, and there's no revenue threshold. If your company runs any hosted-model offering, talk to legal before you ship. I'm not a lawyer; read the full license text linked below.

Detail 2: active parameters set your compute, total parameters set your memory

"52B active" sounds like a mid-size model. It isn't, from a memory point of view.

In an MoE, the router picks 4 of 128 experts per token. Different tokens pick different experts, so every expert has to be loaded and ready. That means:

  • Compute per token scales with the ~49B active parameters. That's why MoE models are cheap to run per token.
  • Memory scales with the ~1T total parameters. You can't load only the "active" ones.

Rough weight-only arithmetic (my estimate, not Mistral's figure):

FP8 : ~1 byte per param x ~1T params = ~1 TB of weights NVFP4 : ~0.5 byte per param x ~1T params = ~0.5 TB of weights (+ scale overhead)

On top of that you need KV cache, and a 1M-token context makes KV cache a serious line item if you actually use long prompts. This is a multi-GPU, multi-node deployment, which is why the Hugging Face example sets both tensor parallel (tp_size) and expert parallel (ep_size) sizes. GGUF quants make local experiments possible, but heavy quantization can cost quality, so re-run your evals on whatever quant you deploy.

How I'd decide: API vs self-host

  1. Start on the API. At $0.70 / $2.10 per 1M tokens, run your real task set first. If the model doesn't win on your evals, the hosting question is moot.
  2. Measure cost per successful task, not per token. Mistral's claim is token efficiency (47% fewer tokens per task than GLM-5.3 on one index), so compare finished tasks, at each reasoning level (Low, Medium, High).
  3. Self-host only for a concrete reason: data residency, air-gapped environments, fine-tuning on data that can't leave, or volume high enough that dedicated GPUs beat per-token pricing. Mistral pitches this as "sovereign AI", and that's the real use case.
  4. Check the license bucket before architecture, not after launch.
  5. Budget for memory, not active params. Plan GPUs around the full weights plus KV cache for your real context lengths.
  6. Keep your harness model-agnostic. An OpenAI-compatible server (vLLM, SGLang) means switching between API and self-hosted is a config change.

The takeaway

Open weights give you options, not a free lunch. Read the license for your exact product shape, size hardware by total parameters, and prove the model wins on your own tasks through the API before you buy a rack of GPUs for it.

Would you self-host a 1T MoE model, or is the API good enough for your use case?

References

原文：https://www.a8q9.cn/en/article/mistral-large-4-is-open-weight-read-the-license-and-do-the-memory-math-before-yo-745092.html

相关文章