Atlas

Models

← All models

Llama 3.1 Tülu 3 405B

Allen Institute for AI (Ai2) · 2025-01-30 · 405.9B parameters

Llama 3.1 Tülu 3 405B is Ai2's final RLVR checkpoint in its 405B Tülu 3 post-training sequence, built from Meta's Llama 3.1 405B base and publicly launched on January 30, 2025. Ai2 applied SFT, DPO, and then RLVR on the MATH training split. The repository retains Llama 3.1's 131,072-token configured limit, although Ai2 notes that its Tülu 3 post-training data were relatively short, with most samples under 2,048 tokens and no long multi-turn data.

Benchmark scores

BenchmarkScore
Artificial Analysis Intelligence Index8.3
GPQA Diamond51.6
Humanity's Last Exam (Text-Only)3.5
LiveCodeBench (Artificial Analysis)29.1
MATH-500 (Artificial Analysis source)77.8
MMLU Pro71.6
SciCode (Artificial Analysis)30.2
Loading Atlas data…