DeepSeek R1 Distill Llama 70B: benchmark results

DeepSeek's official R1 distillation into Llama 3.3 70B Instruct, fine-tuned on R1-generated reasoning traces (January 2025). Provider: DeepSeek. Released 2025-01-20. Access: Open.

Aggregate Rating 1556 ± 8, rank #594 of 1656 rated models, from 544 benchmark results.

Strongest benchmark results

BenchmarkScoreMetricPercentile
AMI-ME - Topic Segmentation0.37WindowDiff segmentation error (0-1) of the model's topic bou100
AbstentionBench - answer unknown - KUQ/Future unknowns - Precision97.25Abstention Precision (%)100
AbstentionBench - false premise - KUQ/False assumptions - Precision96.18Abstention Precision (%)100
AbstentionBench - false premise - QAQA - Precision92.86Abstention Precision (%)100
AbstentionBench - subjective - KUQ/Controversial - Precision96.6Abstention Precision (%)100
AbstentionBench - subjective - MoralChoice - Precision100Abstention Precision (%)100
AbstentionBench - underspecified intent - SituatedQA/Geo - Precision86.21Abstention Precision (%)100
French LLM Leaderboard - Average55.93Mean Normalized Score (%)100
French LLM Leaderboard - GPQA FR63.19Score (%)100
GoTo-AI Leaderboard - IndoMMLU - Social Science71.3Score (0-100)100
Open CoT - LSAT Analytical Reasoning22.61CoT Gain (%)100
ToxReason - Cardiotoxicity78.5F1 (%) for heart toxicity predicted by the model on the ToxR100

Interactive version: theaggregate.ai/model?slug=deepseek-r1-distill-llama-70b · How It Works · Data refreshed daily, snapshot 2026-10-20.