Home›Models›mistral-large-4

Model sheet

mistral-large-4

Architecture mixture of experts72 LayersLicense mistral-ecosystem-open-weightMaximum context 1,024kVocabulary 204,800

Frame from Mistral's Mistral Large 4 launch video: a giant orange pixel “4” in a museum hall, with the “Le Chonk” cat seen from behind
Frame from Mistral's Mistral Large 4 launch video: a giant orange pixel “4” in a museum hall, with the “Le Chonk” cat seen from behind Still from the official video · Mistral AI

Watch the video

At Q4_K_M this is a 591.21 GiB file and it fits on none of the 18 machine profiles we measure at 8k tokens. Of those 18, only the Mac Studio with M5 Ultra and 512 GB runs it, and only at 2 or 3 bits: with the 354.56 GiB Q2_K it reaches, by our count, its full 1024k context.

At a glance
Size at Q4_K_M591.21 GiB
Context memory at 8k0.98 GiB
Fits on0 / 18
German cost×1.24
591.21 GiB

Model size at Q4_K_M

Measured by Local AI Scope

0.98 GiB

Context memory at 8k tokens

Measured by Local AI Scope

×1.24

What German costs versus English

Measured by Local AI Scope

Videos

Videos: mistral-large-4 1

The official videos we found most useful. Press play and it opens right here.

The videos live on YouTube, and nothing loads from there until you press play (no-cookie mode). The thumbnail is served by us.

Sizes

Sizes of the real published files

These are the sizes of the files actually published by the author, added up when the model ships split into parts. Not an estimate from the parameter count.

Size of the published file for each compression
Compression Model size
Q8_0 1,038.57 GiB
Q4_K_M 591.21 GiB

This model also reads images. For that it needs a second file, the vision projector (mmproj-F16.gguf, 3.17 GiB), published next to the model and not included in the figures on this page: they count text only. If you are going to send it images, add it to the model size.

Context

Context memory

Computed from the attention pattern this model declares layer by layer, not from the generic formula. That is why the figure is often far smaller than other calculators tell you. This model stores its context memory compressed (latent attention): per layer and token, one vector of 512 values instead of its 64 heads of 64. We count it as one head of 512 for the key and another for the value, which is a little more than it really takes (one latent serves both), so these figures err on the high side, never the low one.

  • 8k tokens → 0.98 GiB
  • 32k tokens → 3.94 GiB
  • 128k tokens → 15.75 GiB
Languages

What it costs in each language

Measured by running this model’s own tokenizer over the same content in four languages. Tokens are what you pay for in context, in memory and in your cloud bill.

Token cost per kind of text and language, with English as the baseline
Kind of text EN ES FR DE
Code — ×1.18 ×1.27 ×1.37
Contracts — ×1.10 ×1.15 ×1.19
Business email — ×1.11 ×1.17 ×1.13
Support tickets — ×1.17 ×1.31 ×1.25

English is the baseline: every figure is how many times more tokens the same content costs in that language. Average across the whole corpus: DE ×1.24.

Machines

Where it fits

Fits means the model, its context memory and the working headroom all fit at the stated context, leaving the system its share. 0 of the 18 machine profiles tested (Q4_K_M, 8k tokens).

The 18 machine profiles, with the memory each one leaves and the largest context this model holds on it
Machine Fits Memory left Largest context that fits
PC with an RTX 4060 8 GB no 7.20 GiB —
PC with an RTX 3060 12 GB no 11.20 GiB —
PC with an RTX 4070 12 GB no 11.20 GiB —
Mac with M4 and 16 GB unified memory no 13.00 GiB —
Office laptop, CPU only, 16 GB no 13.00 GiB —
Mac mini with M6 and 16 GB no 13.00 GiB —
Mac with M4 Pro and 24 GB no 21.00 GiB —
Workstation with an RTX 4090 24 GB no 23.20 GiB —
Mini-PC with an NPU and 32 GB LPDDR5X no 29.00 GiB —
Mac mini with M6 and 32 GB no 29.00 GiB —
Server with 2× RTX 3090 (48 GB) no 47.20 GiB —
Mac with M4 Max and 64 GB no 61.00 GiB —
Mac mini with M5 Pro and 64 GB no 61.00 GiB —
Xiaomi AI Cube, 80 GB — engineering prototype no 77.00 GiB —
PC with Ryzen AI Max+ 395 and 128 GB unified no 96.00 GiB —
Mac Studio with M5 Max and 128 GB no 125.00 GiB —
PC with Ryzen AI Max+ PRO 495 and 192 GB unified no 160.00 GiB —
Mac Studio with M5 Ultra and 512 GB no 509.00 GiB —

None of the machine profiles we measure takes this model at the reference compression, not even at the smallest context.

Before you hit download: as of 3 November 2026, Mistral calls this model “Le Chonk”, and it is not exaggerating: at 591.21 GiB in Q4_K_M it fits none of our machines. On our list, the only way to run it is to drop to 2 or 3 bits on the Mac Studio with M5 Ultra and 512 GB. By our count, that machine takes ggml-org’s Q2_K (354.56 GiB) with its full context, 1024k; unsloth’s UD-IQ3_XXS (394.43 GiB) up to 512k; and its UD-Q3_K_XL (458.91 GiB) up to 256k. Our catalogue does not list those 2- and 3-bit files, so the Model Finder does not offer them. To open it you need llama.cpp v0.9.0, its stable release from 4 November 2026, or any daily build from b11998 (30 October 2026) onwards. Ollama only offers it in its cloud (mistral-large-4:cloud), and LM Studio has no page for the model yet. mlx-lm does not open it in its latest version, 0.32.2: mlx-community’s MLX versions were made with a change that has not been released yet; the 3-bit one weighs 427.71 GiB (text only), and the 4-bit one (549.86 GiB), its own authors say, does not fit in a single 512 GB Mac. vLLM serves it from its official Docker image, with a tag of its own for this model; Mistral’s recipe uses 8 graphics cards, and its stable release has included it since 0.34.0, published on 16 November 2026. The table counts text only: to send it images, add the vision projector (3.17 GiB). And the licence, the Mistral Ecosystem Open-Weight License, is a modified MIT licence: running it at home or inside your company changes nothing, but if your company (or one in its group) sells third parties access to the model as a service (through an API, for example), Mistral requires a separate commercial licence first.