Mistral AI published the weights of Mistral Large 4 on Hugging Face on 29 October 2026, so anyone can now download it. Very few people can run it at home. The model has about 1 trillion parameters. The smallest version, compressed to roughly 2 bits per parameter, is a 298 GB file and needs a computer with about 512 GB of memory. For everyone else, this news matters mainly because it is a large, permissively licensed open model, not because it will run on your laptop.
What changed on 29 October
Mistral first offered Large 4 as a preview through its own paid API earlier in October. On 29 October it released the full weights. Reports from the launch say the downloadable model has had more training than the preview. Mistral describes it as its best open-weight model for coding.
The basic facts, from Mistral's model card:
| Item | Mistral Large 4 |
|---|---|
| Total parameters | About 1 trillion |
| Active parameters | About 49 billion per token (52 billion counting the embedding and output layers) |
| Model design | Mixture of experts: only part of the model works on each token |
| Published context limit | 1 million tokens |
| Input types | Text and images |
| Official release format | FP8 (8-bit numbers) |
| License | Mistral Ecosystem Open-Weight License, a modified MIT license |
Why "only 49B active" does not make it small
A mixture-of-experts model routes each word through a small group of specialist sub-networks. Only about 49 billion parameters do work on any given token. As a rough rule, once it is loaded, Large 4 generates text at about the speed of an ordinary model of around 50B parameters on the same hardware, not at the speed of a 1-trillion-parameter one.
But every expert must be loaded into memory, because the next token might need any of them. Memory depends on the total parameters, not the active ones. That is the number that rules Large 4 out for ordinary computers. The quantization guide explains how compression shrinks those weights and what it costs.
How much memory each version needs
The llama.cpp project (ggml-org) has published GGUF model files at several compression levels:
| Model file | Download size | What it means |
|---|---|---|
| BF16 | 2.10 TB | 16-bit, a lossless upcast of the official FP8 release |
| Q8_0 | 1.12 TB | 8-bit; close to the original |
| Q4_K_M | 635 GB | 4-bit; the usual balance of size and quality |
| Q2_K | 381 GB | 2-bit; the page says it fits in 512 GB of unified memory |
| IQ2_XXS | 298 GB | The smallest; also listed as fitting in 512 GB of unified memory |
The download size is not the whole requirement. The model also needs conversation memory (KV cache) for whatever you send it, plus room for the model runner itself. That is why the 298 GB file is listed against a 512 GB machine. Using a large part of the 1 million-token context would need more again.
The 2-bit files carry a real cost in answer quality. Mistral's published scores come from the full model, and Local LLM Guide has not measured how much the 2-bit versions lose. For a two-bit conversion, keep its quantization name with your results. Test it on examples you can score and compare with a higher-precision run if available. Mistral's scores describe its tested release and do not quantify the change for this conversion.
What hardware could run it
In plain terms, you need one of these:
- A workstation or server with about 512 GB of memory, for the 2-bit files. On a computer with unified memory (where the processor and graphics share one pool, as on Apple silicon Macs), all of it can hold the model. On a PC, the model would mostly sit in system memory (RAM) and run on the processor, which is much slower than running in graphics memory (VRAM).
- Roughly 640 GB to 700 GB of memory (our estimate, not a published figure), for the 4-bit Q4_K_M file with a modest context length.
- A multi-GPU server, for the official FP8 release. Mistral's examples use eight data-center GPUs running vLLM or SGLang, and its NVFP4 version targets NVIDIA's Blackwell GPUs.
No consumer graphics card comes close. Even a 32 GB card holds only about one-ninth of the smallest file. Splitting the model between a GPU and a large pool of RAM can work in llama.cpp, but speed then depends mostly on how fast the RAM is.
Not sure where your computer stands? What local LLM can my computer run? and the RAM and VRAM calculator give a realistic starting point.
If you have the hardware
The GGUF page gives a direct Ollama command for the 4-bit file:
ollama run hf.co/ggml-org/Mistral-Large-4-1T-A52B-GGUF:Q4_K_M
That command starts a 635 GB download. Check your free disk space and memory before you run it. For the smallest file, the page gives a llama.cpp command instead:
llama serve -hf ggml-org/Mistral-Large-4-1T-A52B-GGUF:IQ2_XXS
Two settings behave differently from most local models:
- Reasoning cannot be turned off. You can set
reasoning_efforttolow,mediumorhigh; the default ishigh. A lower setting gives faster, shorter answers. - Recommended sampling is temperature 1.0 and top_p 0.95, and these are stored in the GGUF files.
Image input needs the separate vision file (mmproj, 3.4 GB) from the same page.
Is Mistral Large 4 open source?
The Mistral Ecosystem Open-Weight License is a modified MIT license. It allows commercial and non-commercial use, modification and fine-tuning, with two conditions. Every copy of the model must keep Mistral's copyright and permission notice. If you plan to offer customers access to the model as a service, read the separate commercial-license requirement before deployment. Internal use that does not make the model available to anyone outside the business is exempt from that condition.
On that wording, a law firm or company running Large 4 on its own servers for its own staff appears to fall within the free terms, while a company selling access to it as a hosted service does not. Mistral itself calls the model "open-weight": you can download and run the weights, under a license with conditions. Read the license itself before a commercial deployment.
What Mistral says about quality
These are Mistral's figures, from its model card. Local LLM Guide has not tested them:
- Coding agents: 48% on the Artificial Analysis Coding Agent Index, which Mistral calls "effectively tied" with GLM-5.3, while using 47% fewer tokens per task on average.
- Overall: 26 on the Artificial Analysis Intelligence Index v5. Mistral's own card concedes that the strongest Chinese open models score higher; Large 4 leads only among Western ones.
- Speed: a built-in draft model (DSpark) for speculative decoding, which Mistral says gives up to 1.9 times higher throughput.
Better choices for most computers
If you want a capable model on hardware you already own, start smaller. The 16 GB RAM setup guide and 32 GB setup guide list models that fit ordinary machines, and local models for coding covers coding-focused options. If you need Large 4 itself but lack the hardware, Mistral's hosted API is the practical route. Text you send there is processed on Mistral's servers, not on your computer, so check the local AI privacy guide before sending confidential material.