DeepSeek R1 Distill Llama 8B: benchmark results
DeepSeek's distilled reasoning model, transferring DeepSeek-R1's reasoning behavior onto a Llama 8B architecture, one of six dense models the team distilled from R1 using Llama and Qwen bases. Provider: DeepSeek. Released 2025-01-20. Access: Open.
Aggregate Rating 1380 ± 11, rank #1291 of 1656 rated models, from 357 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| JailNewsBench (Research Front) - Harmfulness Score | 2.4 | Mean harmfulness of the generated fake news (0 to 4): the av | 100 |
| Open CoT - LogiQA | 13.26 | CoT Gain (%) | 99.2 |
| Open CoT - LogiQA 2 | 18.26 | CoT Gain (%) | 98.5 |
| Open CoT Leaderboard | 16.1 | Average CoT Gain (%) | 96.9 |
| Open CoT - LSAT Reading Comprehension | 22.68 | CoT Gain (%) | 93.9 |
| JailNewsBench (System Override) - Harmfulness Score | 2.5 | Mean harmfulness of the generated fake news (0 to 4): the av | 93.8 |
| RFEval - Paper Review | 94.53 | Contrast-conditional reasoning faithfulness (%): after a cou | 90 |
| Open CoT - LSAT Analytical Reasoning | 8.26 | CoT Gain (%) | 89.3 |
| StakeBench - Commitment Detection | 72.1 | Accuracy (%) on G1: does a comment come from a user holding | 85.7 |
| Open CoT - LSAT Logical Reasoning | 18.04 | CoT Gain (%) | 84.7 |
| JailNewsBench (Context Overload) - Harmfulness Score | 2.1 | Mean harmfulness of the generated fake news (0 to 4): the av | 81.2 |
| FACTS Leaderboard | 40.68 | Combined Score (%) | 75.8 |
Interactive version: theaggregate.ai/model?slug=deepseek-r1-distill-llama-8b · How It Works · Data refreshed daily, snapshot 2026-10-20.