Atlas

Models

← All models

DeepSeek R1 0528 Qwen3 8B

DeepSeek · 2025-05-28 · 8.2B parameters

DeepSeek-R1-0528-Qwen3-8B is an 8B dense reasoning model released May 28, 2025, created by distilling DeepSeek's R1-0528 reasoning model into a Qwen3-8B base backbone. The chain-of-thought distillation transfers strong step-by-step math and reasoning behavior into the smaller open-weight Qwen3 model. Its published configuration extends the Qwen3 backbone's 32K original window to 128K with YaRN scaling.

Benchmark scores

BenchmarkScore
AA-LCR13.0
AA-Omniscience Index-64.1
AIME 202563.7
Artificial Analysis Intelligence Index10.4
Artificial Analysis Omniscience Accuracy12.0
Artificial Analysis Omniscience Hallucination Rate86.5
Artificial Analysis Openness Index47.2
BRIDGE Medical (chain-of-thought)36.3
BRIDGE Medical (few-shot)42.9
BRIDGE Medical (zero-shot)36.9
Capability129.0
CritPt0.0
FormationEval85.1
GPQA Diamond61.2
Humanity's Last Exam (Text-Only)5.6
IFBench (Artificial Analysis)19.9
LiveCodeBench (Artificial Analysis)51.3
MATH-500 (Artificial Analysis source)93.2
SciCode (Artificial Analysis)20.4
SnakeBench Average Apples1.2
SnakeBench Best Apples9.0
SnakeBench Rating23.8
SnakeBench Total Apples48.0
SnakeBench Win Rate56.8
Terminal-Bench Hard1.5
τ²-bench (Telecom)0.0
Loading Atlas data…