What Hardware Can Run Mistral Large 4? Measured Runs and Memory Needs

Short answer

Mistral Large 4 is Mistral’s open-weight model: about 1 trillion parameters, only about 49B used per token. Its smallest file, 233 GB, needs roughly 256 GB or more total memory, beyond any single graphics card or ordinary PC. Measured runs exist on a 512 GB Mac Studio (M3 Ultra, 2-bit file), four RTX Pro 6000 cards and an 8 × H200 server; without hardware, the Mistral API is easiest.

The details

• Measured by someone: a 512 GB Mac Studio (M3 Ultra, ggml-org 2-bit Q2_K file, 15 tok/s, context not stated). Four RTX Pro 6000 cards (2-bit file: 30 tok/s on short prompts, 3 tok/s at long context; NVFP4 file: about 10 tok/s). An 8 × H200 server (FP8, about 65 tok/s for one request, with speculative decoding on).
• Fits on paper, not yet measured (weights-only fit, context unknown): a 512 GB M5 Ultra Mac Studio (2-bit). A 256 GB Mac with its GPU memory limit raised (1-bit only). Several linked DGX Sparks or Macs, where total memory clears the file. Servers with a lot of RAM (the model runs on the CPU). Speed on all of these is Unknown.
• Won’t run it: any single GPU, or any machine with less than about 256 GB of total memory.
• Easiest option: the Mistral API, at $0.70 per million input tokens and $2.10 per million output tokens.
• What we can’t say yet: speed on rigs nobody has measured, and how much memory long context needs. The size of its KV cache (the memory that holds the conversation so far) isn’t known yet.

What the terms mean

• Quant: a smaller copy of the model with each weight stored in fewer bits (2-bit, 4-bit and so on). Fewer bits means a smaller file and some loss of quality.
• Active parameters: the part of the model used for each token. Mistral Large 4 holds about 1T parameters but uses about 49B per token, or 52B counting the embedding and output layers. That is why Mistral’s docs say 52B and some write-ups say 49B.
• Speculative decoding: a small draft model guesses the next few tokens and the big model checks them in one pass. When guesses are accepted it speeds up output, so results with it on are labelled as such. Mistral Large 4 ships with a built-in draft head called DSpark.

Memory needed by file

The third column is our usual fit rule: the file plus 1.5 GB of headroom, before any context. Unsloth’s own guidance adds room for context, so its totals are higher. Sizes are the decimal GB the Hugging Face API reports for each repo, checked Oct 29, 2026.

Sources: Unsloth GGUF repo, ggml-org GGUF repo, official FP8 and official NVFP4 repos, and Unsloth’s Mistral Large 4 guide (read Oct 29, 2026). ggml-org also publishes an NVFP4 GGUF in a separate repo; we couldn’t confirm its size. To send it images, add the vision projector file (mmproj), which takes another 3.4 GB.

On a Mac we count only the GPU’s default share of unified memory, which is 75% above 36 GB (how we count Mac memory). A 512 GB Mac counts as 384 GB, enough for files up to ggml-org Q2_K by our rule, with about 2 GB left for context. A 256 GB Mac counts as 192 GB, which fits nothing. You can raise the limit with sudo sysctl iogpu.wired_limit_mb=<MB>, at the cost of memory for macOS and other apps.

Measured runs

All four runs are from Hugging Face’s Mistral Large 4 blog post, published Oct 29, 2026. That makes them third-party measurements, not Mistral’s. The post doesn’t give runtime versions or say exactly how it timed prefill and generation.

These runs used different files, runtimes, contexts and settings, so don’t compare rows directly. The H200 figures have speculative decoding on and the others don’t say. The blog also says “the 2-bit quants run on a single Apple Silicon Ultra computer with 512 GB of unified RAM” and that the other files need several machines.

Prices for the measured hardware, where we could check them: a used 512 GB M3 Ultra Mac Studio sold for $16,000 to $19,750 on eBay between Sep 10 and Oct 6, 2026 (MacPro-LA). One RTX Pro 6000 Blackwell card was about $17,187 in October 2026 (GPU Poet), not counting the host.

Hardware that fits on paper but has no published run

Usable memory follows our rules: the Mac GPU share above, about 115 GB per 128 GB DGX Spark, and system RAM less 8 GB for the OS. “Fits” means the file plus 1.5 GB fits, with no room checked for context. None of these has a measured speed.

For how linked machines share a model, see our DGX Spark cluster guide and Mac Studio cluster guide.

What won’t run it

• Any single GPU. The largest we track, the 96 GB RTX Pro 6000, holds less than half of the 233 GB 1-bit file.
• One DGX Spark, or two by our rules (see above).
• The 192 GB Framework Desktop, any other single Ryzen AI Max box, and any 128 GB Mac Studio.
• A 256 GB Mac at its default GPU limit (192 GB).

How to check your own rig

1. Add up the memory your rig can use: GPU memory, plus system RAM less 8 GB if you’ll offload, or the GPU share on a Mac. My rig does this for one machine, and the methodology explains each rule.
2. Pick a file from the memory table and add 1.5 GB of headroom. If your total is below that, it won’t load.
3. Leave room for context. We can’t say how much this model needs, so a fit with only a few GB to spare may fail at the context you want.

Our Mistral Large 4 model page checks single devices against the 4-bit file. It doesn’t cover the smaller files or linked machines.

Common pitfalls

• Download size isn’t the memory you need. You need the file plus headroom plus room for context, and a 512 GB Mac only gives the GPU 384 GB by default.
• Long context slows it down sharply. On four RTX Pro 6000s, generation fell from 30 tok/s on short prompts to 3 tok/s at long context.
• Mistral’s official NVFP4 file targets NVIDIA Blackwell GPUs with vLLM or SGLang. It won’t help on a Mac or an older GPU.
• You need a llama.cpp, vLLM or SGLang build with Mistral Large 4 support. The model card lists vLLM (recommended), SGLang, transformers and llama.cpp, and says “we recommend using the Mistral AI API if local serving is subpar”.
• Reasoning can’t be turned off (per Hugging Face’s blog), only set to low, medium or high. Answers include thinking tokens, so a reply takes longer than its visible length suggests.
• Images need the mmproj vision file loaded alongside the model.

Benchmarks (Mistral’s own claims)

These come from Mistral’s open-weights announcement, its post on X and the model card. Mistral also says the model is “effectively tied” with GLM-5.3 on the Coding Agent Index while using 47% fewer tokens per task. Sources disagree on some figures: Hugging Face’s blog gives 49.8% on the Coding Agent Index, 28.3% on Terminal-Bench 4 and 82% on a vulnerability-reproduction part of the Cyber Index, likely from a different snapshot. We haven’t checked these on Artificial Analysis, and we don’t test model quality ourselves.

License in plain words

The weights use the Mistral Ecosystem Open-Weight License, a modified MIT license (LICENSE file). You can use, modify, fine-tune and sell things built on it.

The one condition covers a “Model as a Service” business: giving other people access to the model for inference or further training, with real control over inputs, settings or training data, such as through an API or a hosted endpoint. If you or an affiliate run one, you need a separate commercial license from Mistral before using it commercially. These don’t count as Model as a Service:

• A product for end users with the model built into specific features.
• Passing requests on to a copy of the model that someone else hosts.
• Internal use that isn’t available to anyone outside your company.

Specs

FAQ

Can I run it on one GPU? No. The smallest file is 233 GB and the largest single card we track has 96 GB.

Can a 256 GB Mac run it? Not at the default GPU limit, which counts 192 GB. With the limit raised, the 233 GB 1-bit file fits on paper, but we can’t say how much context fits beside it. Nobody has published a run, so its speed is Unknown.

How can I try it without buying hardware? Use the Mistral API or another hosted provider. You pay $0.70 per million input tokens and $2.10 per million output tokens at Mistral, and you don’t need any hardware.

Can I use it commercially? Yes, including in products you sell. You only need a separate license from Mistral if you run a business that offers the model itself to others as a service, such as a hosted API.