Mistral Large 4 VRAM requirements (estimate)
Mistral Large 4 should need about 652 GB of GPU memory at Q4_K_M with an 8K context and one request: 592 GB of weights plus the KV cache and runtime overhead. That is more than eight H100s hold (640 GB), so the smallest setup on our list is eight H200s (1,128 GB). In FP8 it needs about 1,087 GB, in BF16 2,152 GB. These are estimates: Mistral announced the model on October 6, 2026 with 1.05T total and 49B active parameters, and the weights are due on Hugging Face on October 31, 2026. When they are published, this page switches to figures from the real config.json and files.
652 GB at Q4_K_M, 8K context, one request (estimate)
- FP8, if published like Large 3
- 1,087 GB
- FP16 / BF16
- 2,152 GB
- Weights expected
- Smallest setup, Q4_K_M
- 8× H200 (1,128 GB)
What Mistral has published so far
- Total parameters
- 1.05T (Mistral's docs; the announcement rounds it to 1 trillion)
- Active per token
- 49B, or 52B counting the embeddings and output layer
- Architecture
- Granular mixture of experts, natively multimodal, with a 1.6B vision encoder. The layer count, attention and expert layout come with the weights.
- Context window
- 1M tokens
- Weights
- Expected October 31, 2026 as mistralai/
Mistral-Large-4.0-1T05-A52B ; the repository is not open yet. Until then the model runs only on Mistral's preview API. - License
- Not announced. Mistral Large 3 is Apache 2.0.
- Sources
- Mistral's announcement, the model docs and the Hugging Face page, read .
How much VRAM does Mistral Large 4 need?
Each total is the weights plus an FP16 KV cache for 8,192 tokens and the calculator's runtime overhead (0.5 GB plus 10%), as on every model page. The last two columns are the 512 GB Mac Studio and a PC or server that keeps the experts in system RAM.
| Precision | Weights | Total, 8K | Smallest setup | 512 GB Mac Studio | System RAM, experts offloaded |
|---|---|---|---|---|---|
| FP16 / BF16 | 1,956 GB | 2,152 GB | More than one 8-GPU server | No | 2,048 GB |
| GGUF Q8_0 | 1,039 GB | 1,144 GB | 8× B200 (1,440 GB) | No | 1,536 GB |
| FP8, if published like Large 3 | 987 GB | 1,087 GB | 8× H200 (1,128 GB) | No | 1,024 GB |
| GGUF Q6_K | 802 GB | 883 GB | 8× H200 (1,128 GB) | No | 1,024 GB |
| GGUF Q5_K_M | 693 GB | 763 GB | 8× H200 (1,128 GB) | No | 768 GB |
| GGUF Q4_K_M | 592 GB | 652 GB | 8× H200 (1,128 GB) | No | 768 GB |
| NVFP4, if published like Large 3 | 584 GB | 644 GB | 8× H200 (1,128 GB) | No | 768 GB |
| GGUF IQ4_XS | 523 GB | 576 GB | 8× H100 SXM (640 GB) | No | 768 GB |
| GGUF Q3_K_M | 467 GB | 515 GB | 8× H100 SXM (640 GB) | No | 512 GB |
| GGUF UD-Q2_K_XL | 372 GB | 411 GB | 8× H100 SXM (640 GB) | With the limit at 504 GB | 384 GB |
| GGUF UD-IQ2_XXS | 326 GB | 360 GB | M3 Ultra Mac Studio (512 GB, 384 GB usable) | Yes | 384 GB |
| GGUF UD-IQ1_S | 277 GB | 306 GB | M3 Ultra Mac Studio (512 GB, 384 GB usable) | Yes | 384 GB |
| GGUF UD-TQ1_0 | 245 GB | 270 GB | M3 Ultra Mac Studio (512 GB, 384 GB usable) | Yes | 256 GB |
Q8_0 to Q4_K_M use llama.cpp's average bits per weight, which Mistral Large 3's GGUF files from unsloth match within 0.2% (Q4_K_M: 4.83 bits). IQ4_XS and below vary more between models and uploaders, so they take the bits per weight of those Large 3 files (UD-Q2_K_XL: 3.05). FP8 and NVFP4 scale Mistral's own Large 3 checkpoints, 635 GB in FP8 and 375 GB in NVFP4 for 675B parameters. GB means GiB (1024³ bytes), as everywhere on this site.
Running Mistral Large 4 on a Mac Studio or one GPU
A 512 GB Mac Studio holds the 1- and 2-bit quants. Its GPU may use about 384 GB by default, enough for UD-IQ2_XXS
(360 GB); after sudo sysctl iogpu.wired_limit_mb=516096 it may use 504 GB,
and UD-Q2_K_XL (411 GB) fits too. Q3_K_M (515 GB) does not. How the Mac GPU limit works.
No single GPU comes close. With llama.cpp's --cpu-moe or --n-cpu-moe N the experts stay in system RAM and a
24 or 32 GB card holds the attention layers and the cache. Plan on RAM for the whole file plus a few GB for the OS:
384 GB for UD-Q2_K_XL, 512 GB for Q3_K_M, 768 GB for Q4_K_M. How expert offloading works.
Every token reads the 49B active parameters, as many as DeepSeek V4 Pro reads, so at the same quant on the same hardware the two should write at a similar speed. On eight H200s, FP8 leaves only about 37 GB free for cache, so long contexts or many users call for eight B200s (1,440 GB), which also hold Q8_0 (1,144 GB). IQ4_XS (576 GB) is the largest quant that fits eight H100s.
Long context: the KV cache
Mistral has not published Large 4's attention layout. If it keeps Large 3's, multi-head latent attention over 61 layers caching 576 values per token per layer (params.json), the cache grows by 69 KB per token in FP16. The real figure could differ either way; it is the first number this page will replace.
| Context | FP16 cache | FP8 cache |
|---|---|---|
| 8K | 549 MB | 275 MB |
| 32K | 2.14 GB | 1.07 GB |
| 128K | 8.58 GB | 4.29 GB |
| 1M (full) | 68.6 GB | 34.3 GB |
One request. vLLM and SGLang reserve the cache for many requests at once, so a server sized for heavy use needs several times this. How the KV cache is counted.
Mistral Large 4 next to other big open models
At Q4_K_M with 8,192 tokens of context, the same estimate as each model's own page.
| Model | Parameters | Published files | Q4_K_M total | Smallest setup |
|---|---|---|---|---|
| Mistral Large 3 | 675B, 41B active | 635 GB (FP8) | 419 GB | 8× H100 SXM (640 GB) |
| DeepSeek V3.2 | 685.4B, 37B active | 642 GB | 426 GB | 8× H100 SXM (640 GB) |
| GLM-5.3 | 753.3B, 40B active | 704 GB | 468 GB | 8× H100 SXM (640 GB) |
| MiMo V2.6 Pro | 1.02T, 42B active | 527 GB | 636 GB | 8× H100 SXM (640 GB) |
| Mistral Large 4 (estimate) | 1.05T, 49B active | 987 GB if FP8 | 652 GB | 8× H200 (1,128 GB) |
| DeepSeek V4 Pro | 1.60T, 49B active | 805 GB | 993 GB | 8× H200 (1,128 GB) |
| Kimi K3 | 2.78T, 104B active | 1,454 GB | 1,724 GB | None listed |
What changes when the weights are out
- The weights and the cache come from the published config.json and the real file sizes, replacing every estimate on this page.
- The GGUF files on Hugging Face as they appear, with the one to download for each GPU and Mac, and the llama.cpp and vLLM commands.
- Mistral Large 4 in the calculator's model list, with its GPU and "Can I run" pages.
Mistral Large 4 VRAM questions
When will Mistral Large 4's weights be released?
Mistral says the open weights come at the end of October 2026, and the Hugging Face page of mistralai/
How much VRAM does Mistral Large 4 need?
About 652 GB at Q4_K_M with an 8K context (592 GB of weights), 1,087 GB in FP8 and 2,152 GB in BF16. These are estimates from Mistral's stated 1.05T parameters and the files of Mistral Large 3; this page switches to the real files when they are out.
Can Mistral Large 4 run on a Mac Studio?
On the 512 GB M3 Ultra, only as a 1- or 2-bit quant: UD-IQ2_XXS (about 360 GB) fits the default GPU limit of 384 GB, and UD-Q2_K_XL (about 411 GB) fits after raising the limit to 504 GB. Q3_K_M (515 GB) and larger do not fit.
Can I run Mistral Large 4 on an RTX 4090 or RTX 5090?
Not on the GPU alone. With llama.cpp's --cpu-moe the experts stay in system RAM and the card holds the attention layers and the cache, so you need RAM for the whole file: about 384 GB for UD-Q2_K_XL, 512 GB for Q3_K_M or 768 GB for Q4_K_M.
How many GPUs does Mistral Large 4 need?
Eight H200s (1,128 GB) for Q4_K_M or FP8, eight H100s (640 GB) for Q3_K_M or IQ4_XS, and eight B200s (1,440 GB) for Q8_0. FP8 leaves only about 37 GB free on eight H200s, so long contexts or many users need the B200 server.
Estimated from Mistral's published figures and Mistral Large 3's files, read on . Compare it with other models in the model and GPU report.