Atlas

Models

← All models

DeepSeek-R1-Distill-Qwen-32B

DeepSeek · 2025-01-20 · 32.8B parameters

DeepSeek-R1-Distill-Qwen-32B is a dense reasoning checkpoint fine-tuned from Qwen2.5-32B using approximately 800,000 samples curated with DeepSeek-R1; DeepSeek states that the distilled models used only supervised fine-tuning, without reinforcement learning. Released January 20, 2025, it is the largest Qwen-based R1 distillation, and its released configuration declares a 131,072-token maximum while tokenizer_config.json separately sets model_max_length to 16,384.

Benchmark scores

BenchmarkScore
AA-LCR9.7
Artificial Analysis Intelligence Index11.0
BALROG19.5
BBH (Open LLM Leaderboard v2)42.0
BigCodeBench Complete54.9
BigCodeBench Instruct43.9
BigCodeBench-Hard Complete29.1
BigCodeBench-Hard Instruct23.6
BRIDGE Medical (chain-of-thought)38.7
BRIDGE Medical (few-shot)44.3
BRIDGE Medical (zero-shot)39.8
BRUMO 202568.3
Capability124.8
Creative Writing v273.0
Global PIQA - Non-Parallel (Strict Exact Match)72.2
Global PIQA - Parallel (Strict Exact Match)56.5
GPQA (Open LLM Leaderboard v2)28.4
GPQA Diamond61.5
HMMT Feb 202533.3
Humanity's Last Exam (Text-Only)5.5
IFBench (Artificial Analysis)22.9
IFEval (Open LLM Leaderboard v2)41.9
LiveBench45.5
LiveCodeBench (Artificial Analysis)27.0
MATH Level 5 (Open LLM Leaderboard v2)17.1
MATH-500 (Artificial Analysis source)94.1
MEDIC (clinical summarization)86.2
MEDIC (closed-ended)70.6
MEDIC (open-ended Elo)1503
MMLU Pro73.9
Loading Atlas data…