Atlas

Models

← All models

Mistral Large 3 675B Instruct 2512

Mistral AI · 2025-12-02 · 676B parameters

Mistral Large 3 675B Instruct 2512 is Mistral AI's instruction-tuned FP8 checkpoint of its flagship open-weight multimodal granular Mixture-of-Experts model, released in December 2025 and officially described as having 675B total parameters with 41B active. It is distinct from the Base 2512 checkpoint, accepts text and images, supports a 256K (262,144-token) context window, and is released under the Apache 2.0 license.

Benchmark scores

BenchmarkScore
AA-LCR34.7
AA-Omniscience Index-39.4
AHK-Eval41.7
AHK-Eval Algorithms16.7
AHK-Eval Data Structures16.7
AHK-Eval Date and Time0.0
AHK-Eval Easy50.0
AHK-Eval Hard8.3
AHK-Eval Hidden Cases32.6
AHK-Eval Mid16.7
AHK-Eval Numbers66.7
AHK-Eval Parse Success91.7
AHK-Eval Regex33.3
AHK-Eval Strings16.7
AIME 2024 (avg@8)54.6
AIME 2024-202542.9
AIME 202538.0
AIME 2025 (avg@8)31.3
ALE-Bench721
Antidote: Everyday Edition947
Artificial Analysis Agentic Index5.5
Artificial Analysis Coding Index20.1
Artificial Analysis Intelligence Index15.9
Artificial Analysis Omniscience Accuracy24.1
Artificial Analysis Omniscience Hallucination Rate83.7
Artificial Analysis Openness Index38.9
BullshitBench v2 Leaderboard2.0
BuseyBench SVG3.2
Buyout Game Bradley-Terry Rating1206
Capability134.3
Loading Atlas data…