How much VRAM to run Mistral Large 3?

About 455 GB atQ4_K_M with an 8K context — beyond any single common device. Full breakdown below, or check your exact hardware.

Mistral Large 3 VRAM by quantisation

QuantisationWeightsTotal (8K ctx)Fits on
Q4_K_M411.8 GB454.6 GBmulti-GPU / datacenter only
Q5_K_M480.6 GB530.3 GBmulti-GPU / datacenter only
Q6_K553.5 GB610.5 GBmulti-GPU / datacenter only
FP8 (native)675.0 GB744.1 GBmulti-GPU / datacenter only

Check your hardware

About Mistral Large 3

Mistral Large 3 is the flagship of Mistral's open-weight range: an Apache 2.0-licensed, granular mixture-of-experts model with a vision encoder built in, which its model card sums up as a state-of-the-art general-purpose multimodal system. The active-parameter slice per token is a small fraction of the very large total, and latent-compression attention keeps its long context affordable in memory.

Mistral pitches it at long-document understanding, daily-driver assistants, agentic work and coding — flagship duties — with the unusual property that the same weights are downloadable, so nothing about it is locked to Mistral's own API.

Mistral Large 3 is Mistral's 675B-parameter model released in December 2025, with a 256K-token context window. It is a mixture-of-experts model: all 675B parameters must sit in memory, but only ~41B are active per token, which is what makes it fast for its size. It uses a latent-attention (MLA) design that compresses its KV cache, keeping even very long contexts cheap in memory: its KV cache is about 0.6 GB at an 8K context, 9.2 GB at 128K, and 18.4 GB at the full 256K window (FP16 cache).

Local deployment is datacenter-shaped: Mistral's own guidance points at a node of current-generation accelerators for the native FP8 release, and the totals above show why — the full expert pool must be resident, putting even aggressive quantisation beyond single consumer cards. Two design choices soften the picture at the margins: FP8-native weights mean the primary download is already half the size of a classic full-precision release, and the compressed-attention cache barely grows with context, so long-document work adds little on top. Realistically this page is for capacity planning on shared infrastructure, not desktop ambitions.

Mistral ships FP8-quantised weights as the primary release; community GGUF quants (Q4–Q6) are what most people run locally. Totals above include the KV cache and a realistic framework overhead, so they are what you should expect to see in practice rather than just the download size. Weight sizes are calibrated against real GGUF files — see the methodology.

Frequently asked questions

How much VRAM does Mistral Large 3 need?

At Q4_K_M with an 8K context, Mistral Large 3 needs about 455 GB (weights 412 GB + KV cache + overhead). No single consumer or workstation device fits it — it needs multi-GPU or datacenter hardware.

Can an RTX 4090 (24GB) run Mistral Large 3?

Not fully in VRAM. Mistral Large 3 needs about 455 GB even at Q4_K_M, so a 24 GB card would have to offload layers to system RAM at a large speed penalty.

Can a Mac run Mistral Large 3?

Not on any current configuration — even a 512 GB Mac Studio's ~384 GB GPU allowance is below the ~455 GB this model needs.

How much VRAM does Mistral Large 3 need at its full 256K context?

The KV cache grows with context: about 0.6 GB at 8K tokens versus 18.4 GB at the full 256K window (FP16 cache). Add that to the weights (412 GB at Q4_K_M) plus overhead — long contexts can cost more than a whole quantisation step.

Related

VRAM calculator for any model ·Token counter · Mistral Large 3 API pricing

Similar-size local models: GLM-5.2 (753B) · Llama 4 Maverick (400B) · Qwen3.5 397B-A17B (397B) · DeepSeek V4 Flash (284B) · Mistral Small 4 (119B)

Last updated 2026-09-03. Architecture figures from the model's published config.json; see the methodology.