Atlas

Models

← All models

DeepSeek R1 Distill Llama 8B

DeepSeek · 2025-01-20 · 8B parameters

DeepSeek-R1-Distill-Llama-8B is a dense, text-only reasoning checkpoint released by DeepSeek on January 20, 2025. It was created by supervised fine-tuning Llama-3.1-8B Base on about 800,000 samples curated with DeepSeek-R1, without an RL stage; the downloadable checkpoint contains exactly 8,030,261,248 parameters. Its model config declares 131,072 maximum positions, while its tokenizer config declares a 16,384-token model maximum.

Benchmark scores

BenchmarkScore
AA-LCR0.0
AIME 202541.3
Artificial Analysis Intelligence Index6.4
BBH (Open LLM Leaderboard v2)32.4
BigCodeBench Complete15.3
BigCodeBench Instruct10.6
BigCodeBench-Hard Complete3.4
BigCodeBench-Hard Instruct2.0
BRIDGE Medical (chain-of-thought)27.3
BRIDGE Medical (few-shot)32.9
BRIDGE Medical (zero-shot)28.5
Capability107.2
GPQA (Open LLM Leaderboard v2)25.5
GPQA Diamond30.2
Humanity's Last Exam (Text-Only)4.2
IFBench (Artificial Analysis)17.6
IFEval (Open LLM Leaderboard v2)37.8
LiveCodeBench (Artificial Analysis)23.3
MATH Level 5 (Open LLM Leaderboard v2)22.0
MATH-500 (Artificial Analysis source)85.3
MEDIC (clinical summarization)85.8
MEDIC (closed-ended)48.4
MEDIC (open-ended Elo)1421
MMLU Pro54.3
MMLU-Pro (Open LLM Leaderboard v2)20.9
MuSR (Open LLM Leaderboard v2)32.5
SciCode (Artificial Analysis)11.9
Loading Atlas data…