Mistral Large 4 hardware requirements: VRAM, GPUs & quantization
- Page updated
- Sources checked
- File sizes from Mistral's and the GGUF makers' Hugging Face repos
- Device counts are labeled capacity estimates
- Jump to the checker ↓
Mistral Large 4 ("le Chonk", Mistral's own nickname for it) is a 1.05-trillion-parameter mixture-of-experts model with 52B parameters active per token (49B not counting embeddings and output layers). Mistral released it as a public preview on October 6, 2026 and published the open weights on Hugging Face on October 29, 2026. Here is how much memory it takes to run it yourself, with the real file sizes.
Short answer: Mistral's official FP8 files are 1,060.6 GB (1.06 TB) and its NVFP4 files 610.0 GB. Community GGUFs range from 233.0 GB (Unsloth's 1-bit UD-IQ1_S) and 297.7–380.7 GB at 2-bit to 634.8 GB at Q4_K_M. Only 52B parameters are active per token (49B not counting embeddings and output layers), but all of them must be held in memory. By capacity that is an 8-GPU server at FP8, a 512 GB Mac Studio for the 1–2-bit GGUFs, or a PC with roughly 256 GB of RAM plus GPU memory for the smallest one; it will not fit an RTX 4090/5090 alone. llama.cpp added support on its main branch on October 29; Ollama has only a cloud tag.
How much VRAM does Mistral Large 4 need?
File sizes are the bytes Hugging Face lists for each format's files. Two formats have no file of that name yet, so their size is our estimate from the parameter count (parameters × assumed bits per weight ÷ 8). Device counts add 10% for KV cache and buffers and assume a usable budget per device (H200 131 GB, B200 168 GB, H100 74 GB, 512 GB Mac 460 GB). They are theoretical capacity minimums, not validated deployments.
| Format | Size | Files | H200 | B200 | H100 | Mac 512 GB |
|---|---|---|---|---|---|---|
| BF16 | 2,098.9 GB | ggml-org GGUF | 18 | 14 | 32 | 6 |
| FP8 | 1,060.6 GB | Mistral (official) | 9 | 7 | 16 | 3 |
| Q8_0 | 1,115.2 GB | ggml-org GGUF | 10 | 8 | 17 | 3 |
| Q6_K | ~866 GB | RigCheck estimate (6.6 bits) | 8 | 6 | 13 | 3 |
| Q4_K_M | 634.8 GB | ggml-org GGUF | 6 | 5 | 10 | 2 |
| NVFP4 | 610.0 GB | Mistral (official) | 6 | 4 | 10 | 2 |
| MXFP4 | 577.9 GB | AMD (Quark) | 5 | 4 | 9 | 2 |
| Q3_K_M | ~512 GB | RigCheck estimate (3.9 bits) | 5 | 4 | 8 | 2 |
| Q2_K | 380.7 GB | ggml-org GGUF | 4 | 3 | 6 | 1 |
| IQ2_XXS | 297.7 GB | ggml-org GGUF | 3 | 2 | 5 | 1 |
| UD-IQ1_S | 233.0 GB | Unsloth GGUF | 2 | 2 | 4 | 1 |
The official FP8 and NVFP4 files include the 1.6B vision encoder and the DSpark speculative-decoding draft head; the GGUF sizes exclude the vision projector, a separate 3.4 GB file. Unsloth also publishes UD-IQ1_M (259.6 GB), UD-IQ2_XXS (302.5 GB), UD-Q2_K_XL (378.3 GB), UD-IQ3_XXS (423.5 GB), UD-Q3_K_XL (492.7 GB) and UD-Q4_K_XL (604.8 GB) through UD-Q6_K_XL (890.5 GB). Unsloth reports that its UD-IQ1_S keeps 90.6% top-1 agreement with the full model and UD-Q2_K_XL 95.6%. The Q6_K and Q3_K_M estimates use an assumed whole-model average above the block rate in llama.cpp ggml-common.h; set your own in the calculator.
Mistral Large 4 VRAM & memory calculator
Open in a full page ↗Loading private checker…
- Developer
- Mistral AI Official
- Total parameters
- 1.05 trillion Official
- Active parameters
- 52B per token Official
- Context
- 1M tokens Official
- Official files
- official FP8 files are 1,060.6 GB Official
- Smallest GGUF
- 233.0 GB for Unsloth's UD-IQ1_S Community
- License
- Mistral Ecosystem Open-Weight License Official
- llama.cpp
- merged support for the model on October 29, 2026 Official
- Ollama
- only the cloud tag mistral-large-4:cloud Official
Can you run Mistral Large 4 on a consumer GPU?
Not on GPUs alone: even the smallest file, UD-IQ1_S (233.0 GB), is more than seven RTX 5090s. The home-scale options by memory capacity:
- One 512 GB Mac Studio (M3 Ultra): room for Q2_K (380.7 GB), IQ2_XXS (297.7 GB) or UD-IQ1_S under our assumptions; ggml-org's model card says its Q2_K and IQ2_XXS fit in 512 GB of unified memory. Hugging Face reports about 90 tokens/s prefill and 15 tokens/s generation for Q2_K on one M3 Ultra reported by Hugging Face. Q4_K_M (634.8 GB) needs two clustered machines, via software such as exo or llama.cpp RPC.
- A 256 GB Mac Studio or a PC with about 256 GB of RAM plus GPU memory: only UD-IQ1_S. Unsloth says it "should fit on 256GB RAM including a Mac Studio", and its guide lists 256–270 GB of total memory for its 1-bit files. That is tight: our calculator's 256 GB Mac budget (225 GB usable, plus 10% for the cache) says it does not fit, and the finder gives the same answer for any 256 GB Mac, so treat it as untested. For a PC the finder only adds VRAM and RAM and keeps 10% free, with no separate reserve for the operating system, so a 256 GB RAM build with a graphics card gets a "with caveats" answer there; it is just as untested.
- A server with 512 GB–1 TB of RAM and one large GPU: llama.cpp-style MoE offload keeps shared layers on the GPU and experts in RAM. The idealized batch-one estimate on 8-channel DDR5 (~300 GB/s) is ~9.5 tokens/s at Q4_K_M and ~16 at Q2_K under those assumptions; GPU-resident layers, batching or speculative decoding change it.
Mixture-of-experts helps speed, not memory: only 52B parameters work per token, but the router can pick any expert, so all of them must be loaded. For workstation cards, Hugging Face reports about 30 tokens/s generation for IQ2_XXS at short context on 4× RTX PRO 6000, falling to about 3 tokens/s at long context, and about 10 tokens/s for NVFP4 at 32K, a file larger than the four cards' 384 GB combined, so partly in system RAM (Hugging Face does not give the split) reported by Hugging Face.
Self-host vs. the Mistral API
The comparison is inside the private calculator. Open the cost calculator →.
Mistral Large 4 specs
| Total parameters | 1.05 trillion (Mistral docs; the model card rounds it to 1 trillion). The docs also list a 1.6B vision encoder; the official files include it, the GGUFs keep it in a separate 3.4 GB file. Our estimates for formats without a file use 1.05T. |
|---|---|
| Active parameters | 52B per token (mixture of experts; Mistral docs and announcement). Mistral's Hugging Face page splits it: 49B per token, 52B including embeddings and output layers. Our speed estimate uses 52B. |
| Architecture | 72 layers; 128 routed experts (4 active per token) plus 1 shared expert; Multi-head Latent Attention with a sparse-attention indexer that picks up to 2,048 earlier positions; KV state shared across some layers (model card, Hugging Face blog) |
| Type | Hybrid instruct + reasoning (low, medium, high effort; reasoning cannot be turned off, per ggml-org's GGUF model card), natively multimodal (text and images in, text out) |
| Languages | 160+ |
| Nicknames | ML4, le Chonk (both used in Mistral's announcement) |
| Release | Public preview October 6, 2026; open weights on Hugging Face October 29, 2026 (mistralai/Mistral-Large-4-1T-A52B, FP8, and -NVFP4) |
| Context length | 1M tokens (1,048,576 in the model's config) |
| License | Mistral Ecosystem Open-Weight License: a modified MIT license for commercial and non-commercial use; a business that operates a Model-as-a-Service offering needs a separate commercial license from Mistral, internal use does not |
| Runtimes | vLLM (recommended by Mistral) and SGLang with 8-GPU examples; Transformers; llama.cpp on its main branch since October 29, 2026 (no numbered release when checked October 30; the Llama App and Unsloth Desktop also run the GGUFs, per their makers); Ollama only as a cloud model |
| API | mistral-large-4, $0.70 / $2.10 per million input / output tokens (from October 29, 2026) |
Common setups
The checker's own answer for 8 common setups (graphics cards with 32 or 64 GB of system RAM, then Macs), in the finder's quick mode (graphics cards: the largest file in the Mistral Large 4 calculator's list (official, ggml-org, AMD and Unsloth's UD-IQ1_S) that leaves 10% of VRAM + RAM free, else the smallest if it fits at all; Macs: the largest file in the Mistral Large 4 calculator's list that fits the calculator's usable budget for a Mac of that memory, with 10% for the cache). Answers are capacity estimates, not measured setups; "Unknown" means the checker has no route or source for that hardware. Your own setup: the checker.
| Setup | Answer |
|---|---|
| GeForce RTX 3060 (12 GB) + 32 GB RAM | ✕ Does not fit Even the smallest file, UD-IQ1_S (233 GB), is larger than this setup's 44 GB of memory (VRAM + system RAM). |
| GeForce RTX 4060 Ti (16 GB) + 32 GB RAM | ✕ Does not fit Even the smallest file, UD-IQ1_S (233 GB), is larger than this setup's 48 GB of memory (VRAM + system RAM). |
| GeForce RTX 3090 / 3090 Ti (24 GB) + 64 GB RAM | ✕ Does not fit Even the smallest file, UD-IQ1_S (233 GB), is larger than this setup's 88 GB of memory (VRAM + system RAM). |
| GeForce RTX 5090 (32 GB) + 64 GB RAM | ✕ Does not fit Even the smallest file, UD-IQ1_S (233 GB), is larger than this setup's 96 GB of memory (VRAM + system RAM). |
| Apple Silicon Mac (64 GB unified) | ✕ Does not fit Even the smallest file, UD-IQ1_S (233 GB), is larger than this setup's 64 GB of memory. |
| Mac with M5 Max, 40-core GPU (128 GB unified) | ✕ Does not fit Even the smallest file, UD-IQ1_S (233 GB), is larger than this setup's 128 GB of memory. |
| Mac Studio M3 Ultra (256 GB unified) | ✕ Does not fit Even the smallest file, UD-IQ1_S (233 GB), plus 10% for the cache is more than RigCheck's usable budget for this Mac (225 GB of 256 GB). Unsloth says its 1-bit file should fit on 256 GB; RigCheck's calculator says it does not, so treat it as untested. |
| Mac Studio M3 Ultra (512 GB unified) | ! Fits with caveats: Q2_K (380.7 GB), unified memory (after raising the Mac's GPU wired-memory limit) (capacity estimate) Fits RigCheck's usable budget for this Mac (460 GB of 512 GB) with 10% for the KV cache and buffers. llama.cpp added Mistral Large 4 on its main branch on October 29, 2026; no numbered release included it when RigCheck checked (October 30): use a build of its main branch, or the Llama App (Hugging Face) or Unsloth Desktop that the GGUF makers name. RigCheck capacity estimate from the real file size, not a measured setup. |
Machine-readable: the same answers as JSON (CC BY 4.0, attribution "RigCheck").
Related hardware
Graphics card and Mac pages where this page's checker gives an answer, smallest memory first: graphics cards with 64 GB of system RAM, Macs with their unified memory, in the finder's quick mode (capacity estimate; each page lists every model).
- Mac Studio M3 Ultra (512 GB): fits with caveats: Q2_K (380.7 GB), unified memory (after raising the Mac's GPU wired-memory limit) (Fits RigCheck's usable budget for this Mac (460 GB of 512 GB) with 10% for the KV cache and buffers. llama.cpp added Mistral Large 4 on its main branch on October 29, 2026; no numbered release included it when RigCheck checked (October 30): use a build of its main branch, or the Llama App (Hugging Face) or Unsloth Desktop that the GGUF makers name. RigCheck capacity estimate from the real file size, not a measured setup)
All graphics cards · All Macs · Any other machine
FAQ
How much VRAM does Mistral Large 4 need?
The official FP8 files are 1,060.6 GB and the official NVFP4 files 610.0 GB, plus memory for the KV cache. At FP8 that is the capacity of an 8-GPU server such as 8× B200 or 8× MI300X; Mistral's own vLLM and SGLang examples use 8 GPUs. Community GGUFs need less: 634.8 GB at Q4_K_M, 380.7 GB at Q2_K and 233.0 GB for Unsloth's UD-IQ1_S.
Can Mistral Large 4 run on an RTX 4090 or RTX 5090?
Not on the card alone. Even the smallest GGUF, Unsloth's UD-IQ1_S, is 233.0 GB, far beyond one 24–32 GB card. A PC whose RAM and GPU memory add up to roughly 256 GB or more can hold it by capacity with llama.cpp keeping most experts in system RAM, with speed limited by RAM bandwidth; we know of no measurement of that setup.
Can I run Mistral Large 4 with a 24 GB GPU and 128 GB of RAM?
No. 24 GB of VRAM plus 128 GB of RAM is 152 GB, and the smallest published file, Unsloth's UD-IQ1_S, is 233.0 GB before the KV cache. Offloading the experts to system RAM does not shrink them: the router can pick any expert, so they all have to fit somewhere. Setups that stream experts from an SSD are untested; we know of no measurement. On such a PC, Mistral's API or Ollama's cloud model is the way to use it. A dense model that does fit there is Mistral Medium 3.5 (128B): by our estimate its 4-bit GGUF runs on a 24 GB GPU with most of its layers in system RAM, slowly (the Medium 3.5 checker shows the split for your RAM). See the Mistral Medium 3.5 checker.
Can I run Mistral Large 4 on a Mac Studio?
A 512 GB M3 Ultra can hold the Q2_K (380.7 GB), IQ2_XXS (297.7 GB) or UD-IQ1_S (233.0 GB) GGUF, after raising the GPU memory limit; Hugging Face reports about 15 tokens/s generation for Q2_K on one M3 Ultra. Q4_K_M needs two clustered machines. On a 256 GB Mac only UD-IQ1_S comes close: Unsloth says it should fit, our budget says it is too tight. The GGUFs need a llama.cpp build with Mistral Large 4 support. Community MLX conversions (text only) are also on Hugging Face; RigCheck has not checked MLX support for this model.
Does llama.cpp support Mistral Large 4?
On its main branch, yes: llama.cpp merged support for the model on October 29, 2026 (pull request #31376, "model: mistral le chonk"), and the llama.cpp team (ggml-org) and Unsloth published GGUF files the same day. When we checked on October 30, no numbered llama.cpp release included it yet, so use a build of the main branch, or the tools the GGUF makers name: Hugging Face's Llama App (llama serve, in ggml-org's model card and the Hugging Face blog) or Unsloth's own llama.cpp build and Unsloth Desktop.
Why does a 52B-active model need memory for 1.05T parameters?
In a mixture-of-experts model a router picks different experts for every token, so all experts must be loaded. Active parameters decide speed (bytes read per token); total parameters decide memory.
What is Le Chonk (Mistral Chonk)?
Le Chonk is Mistral AI's own nickname for Mistral Large 4: the announcement says "Unofficially ML4, very officially: le Chonk." The model card is titled "Mistral Large 4 - Le Chonk". It is the same 1.05T-parameter model with 52B active parameters, so every memory figure on this page applies to it.
When were the Mistral Large 4 weights released?
On October 29, 2026, on Hugging Face as mistralai/Mistral-Large-4-1T-A52B (FP8) and mistralai/Mistral-Large-4-1T-A52B-NVFP4. The earlier placeholder name mistralai/Mistral-Large-4.0-1T05-A52B now redirects there (checked October 30, 2026). Mistral had announced the weights for "by the end of the month" (October 2026).
What license does Mistral Large 4 use?
The Mistral Ecosystem Open-Weight License, a modified MIT license that allows commercial and non-commercial use. One condition: a business that operates a Model-as-a-Service offering must get a separate commercial license from Mistral before using the model commercially; internal use and use through Mistral's own products are exempt.
Is Mistral Large 4 on Ollama?
Only as a cloud model so far. Ollama's library lists mistral-large-4:cloud (checked October 30, 2026), which runs on Ollama's servers, not on your hardware, so none of the memory figures on this page apply to it. Ollama lists no local tags yet, even though the weights are out; whether Ollama will offer a local build is not known until it ships one. llama.cpp's main branch and Mistral's own API are the other ways to use it today.
Sources
- Hugging Face: mistralai/Mistral-Large-4-1T-A52B (model card, FP8 files 1,060.6 GB, config, license, vLLM and SGLang examples; read October 30, 2026)
- Mistral docs: Mistral Large 4 model page (1.05T total and 52B active parameters, 1.6B vision encoder, context, API price)
- Mistral AI: Mistral Large 4 is now open weight (license terms, API price from October 29, 2026)
- Hugging Face: mistralai/Mistral-Large-4-1T-A52B-NVFP4 (official NVFP4 files, 610.0 GB)
- Hugging Face: ggml-org/Mistral-Large-4-1T-A52B-GGUF (BF16, Q8_0, Q4_K_M, Q2_K, IQ2_XXS and vision projector files)
- Hugging Face: unsloth/Mistral-Large-4-1T-A52B-GGUF (UD-IQ1_S 233.0 GB and the other Unsloth Dynamic files)
- Unsloth: Mistral Large 4 run-locally guide (memory table by bit width, top-1 agreement, llama.cpp commands)
- Hugging Face: amd/Mistral-Large-4-1T-A52B-Quark-MXFP4 (MXFP4 files, 577.9 GB)
- llama.cpp pull request #31376: model: mistral le chonk (merged October 29, 2026; tags checked October 30, 2026)
- Ollama library: mistral-large-4 tags (only the cloud tag
mistral-large-4:cloud; checked October 30, 2026) - Hugging Face blog: Welcome Le Chonk (architecture, llama.cpp quants, M3 Ultra and RTX PRO 6000 speeds)
- Mistral AI: Mistral Large 4 announcement (parameters, release plan, the "le Chonk" nickname)
- MLX LM README (raising the macOS GPU memory limit)
- Apple: Mac Studio (2025) tech specs (819 GB/s, 512 GB)
- Device budgets, overhead and the Q6_K / Q3_K_M averages are RigCheck assumptions, stated above.