The Open-Source AI Stack
RSS
All models

Models · mistral-large

Mistral Large 3

Open weights Mistral AI · 2025-12-02 · Apache-2.0

Mistral's first MoE since the Mixtral series, released December 2 2025 as base and instruct under Apache 2.0. 675B total / 41B active, trained from scratch on 3000 NVIDIA H200 GPUs. Multimodal with a 256K context, distributed day one on Amazon Bedrock, Azure Foundry, and Hugging Face.

Cost

$0.50 / Mtok input
$1.50 / Mtok output

· as of 2026-05-21

source ↗

Speed

54.2 tok/sec output
643 ms TTFT

· as of 2026-05-21

source ↗

Architecture

tokens in Embedding vocab not disclosed · mistral tokenizer × N layers Attention (not disclosed) Position encoding not disclosed context 256,000 tokens MoE Router ? experts total · ? active per token Output projection tokens out
Schema-generated from data/models.yaml. Every label is auditable against the model's sources.

Specs

Architecture
moe
Total params
675B
Active params
41B
Context window
256K tokens
Attention
unknown
Position encoding
unknown
Training hardware
H200
Post-training
sft, rlhf
OSI-approved
yes
Data released
no
Training code
not released

Benchmarks

Each score carries the date it was published; we never infer or interpolate missing scores.

General reasoning

MMLU-Pro 80.7 as of 2026-05-21 source ↗
GPQA-Diamond 68.0 as of 2026-05-21 source ↗

Code

LiveCodeBench 46.5 as of 2026-05-21 source ↗

Math

AIME 2025 38.0 as of 2026-05-21 source ↗

Available quantizations

GGUF llama.cpp's container; the common local format, k-quants from Q2 to Q8. runs on llama.cpp, Ollama

Verified via the Hugging Face model tree ↗. Community quantizations change over time; the families shown are those with published weights at audit time.

Notable innovations

  • · Mistral's first MoE since Mixtral
  • · Trained on 3000 H200 GPUs from scratch
  • · Apache 2.0 at 675B scale
  • · Native image understanding

Lineage

First Mistral MoE since the Mixtral series; 675B / 41B active under Apache 2.0.

Derived from

Mistral Large 2 2024-07-24

Sources