Atlas

Models

← All models

DeepSeek R1 Distill Qwen 14B

DeepSeek · 2025-01-20 · 14.8B parameters

DeepSeek-R1-Distill-Qwen-14B is a dense reasoning checkpoint produced by supervised fine-tuning of Qwen2.5-14B on 800k samples curated with DeepSeek-R1, without an RL stage. Released January 20, 2025 under the MIT license, its architecture config declares 131,072 positions, while the distillation fine-tuning used a maximum sequence length of 32,768 tokens.

Benchmark scores

BenchmarkScore
AA-LCR7.0
Artificial Analysis Intelligence Index9.8
BBH (Open LLM Leaderboard v2)59.1
BigCodeBench Complete48.4
BigCodeBench Instruct38.1
BigCodeBench-Hard Complete20.9
BigCodeBench-Hard Instruct20.9
BRIDGE Medical (chain-of-thought)34.8
BRIDGE Medical (few-shot)41.4
BRIDGE Medical (zero-shot)34.3
BRUMO 202568.3
Capability125.5
Creative Writing v255.0
Global PIQA - Non-Parallel (Strict Exact Match)70.7
Global PIQA - Parallel (Strict Exact Match)51.8
GPQA (Open LLM Leaderboard v2)38.8
HMMT Feb 202531.7
Humanity's Last Exam (Text-Only)4.4
IFBench (Artificial Analysis)22.1
IFEval (Open LLM Leaderboard v2)43.8
LiveCodeBench (Artificial Analysis)37.6
MATH Level 587.1
MATH Level 5 (Open LLM Leaderboard v2)57.0
MATH-500 (Artificial Analysis source)94.9
MEDIC (clinical summarization)85.8
MEDIC (closed-ended)65.0
MEDIC (open-ended Elo)1501
MMLU Pro74.0
MMLU-Pro (Open LLM Leaderboard v2)46.7
MuSR (Open LLM Leaderboard v2)53.7
Loading Atlas data…