InternLM3-8B-Instruct: benchmark results
Shanghai AI Lab's InternLM3-8B-Instruct, an 8B general-purpose instruction model trained on 4T tokens, with a switchable deep-thinking (long chain-of-thought) mode alongside a normal response mode. Provider: Shanghai AI Lab. Released 2025-01-15. Access: Open.
Aggregate Rating 1477 ± 14, rank #852 of 1635 rated models, from 149 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| FinSafetyBench (FlipAttack) - Financial Crimes (Chinese) | 36.67 | Attack success rate (%; FlipAttack jailbreak; share of the b | 100 |
| FinSafetyBench (FlipAttack) - Ethical Violations (English) | 18.17 | Attack success rate (%; FlipAttack jailbreak; share of the b | 83.3 |
| FinSafetyBench (FlipAttack) - Ethical Violations (Chinese) | 32.63 | Attack success rate (%; FlipAttack jailbreak; share of the b | 80 |
| FinSafetyBench (PAIR) - Ethical Violations (Chinese) | 63.9 | Attack success rate (%; PAIR jailbreak; share of the benchma | 80 |
| FinSafetyBench (PAIR) - Financial Crimes (Chinese) | 80.73 | Attack success rate (%; PAIR jailbreak; share of the benchma | 80 |
| FinSafetyBench - Ethical Violations (Chinese) | 46.92 | Attack success rate (%; mean over the PAIR, ReNeLLM and Flip | 80 |
| FinSafetyBench - Financial Crimes (Chinese) | 64.96 | Attack success rate (%; mean over the PAIR, ReNeLLM and Flip | 80 |
| OpenCompass Language - NLP - Chinese (CompassBench 2501) | 63.5 | Score (%) | 79.5 |
| OpenCompass Language - Dialogue - English (CompassBench 2501) | 63.3 | Score (%) | 74.4 |
| CASTLE Student Safety (Chinese) - Academic Integrity and Competence | 2 | Average Safety Score (1-5) on CASTLE's Chinese scenarios, th | 70.6 |
| CASTLE Student Safety (Chinese) - Excessive Seeking Simplification | 1.92 | Average Safety Score (1-5) on CASTLE's Chinese scenarios, th | 70.6 |
| OpenCompass Language - Dialogue - Chinese (CompassBench 2501) | 69.1 | Score (%) | 69.2 |
Interactive version: theaggregate.ai/model?slug=internlm3-8b-instruct · How It Works · Data refreshed daily, snapshot 2026-10-15.