Atlas

Models

← All models

DeepSeek-R1-Distill-Llama-70B

DeepSeek · 2025-01-20 · 70.6B parameters

DeepSeek-R1-Distill-Llama-70B is a dense reasoning checkpoint fine-tuned from Llama-3.3-70B-Instruct using reasoning data generated by DeepSeek-R1 and released January 20, 2025. Its released configuration declares a 131,072-token maximum via Llama 3 RoPE scaling from an original 8,192-position setting, while tokenizer_config.json separately sets model_max_length to 16,384.

Benchmark scores

BenchmarkScore
AA-LCR11.0
AA-Omniscience Index-46.5
Artificial Analysis Intelligence Index9.9
Artificial Analysis Omniscience Accuracy19.3
Artificial Analysis Omniscience Hallucination Rate81.5
Artificial Analysis Openness Index36.1
BBH (Open LLM Leaderboard v2)56.3
BigCodeBench Complete49.9
BigCodeBench Instruct35.3
BigCodeBench-Hard Complete20.3
BigCodeBench-Hard Instruct20.3
BRIDGE Medical (chain-of-thought)39.0
BRIDGE Medical (few-shot)46.2
BRIDGE Medical (zero-shot)39.8
BRUMO 202566.7
BuseyBench SVG3.0
Capability127.5
Creative Writing v276.2
CritPt0.0
GPQA (Open LLM Leaderboard v2)26.5
HMMT Feb 202533.3
Humanity's Last Exam (Text-Only)6.1
IFBench (Artificial Analysis)27.6
IFEval (Open LLM Leaderboard v2)43.4
LiveBench54.5
LiveCodeBench (Artificial Analysis)26.6
LiveCodeBench Pro293
MATH Level 589.9
MATH Level 5 (Open LLM Leaderboard v2)30.7
MATH-500 (Artificial Analysis source)93.5
Loading Atlas data…