DeepSeek-V2-Lite-Chat: benchmark results
DeepSeek's DeepSeek-V2-Lite-Chat, a 16B-total/2.4B-active MoE chat model using Multi-head Latent Attention and DeepSeekMoE, trained from scratch on 5.7T tokens, deployable on a single 40GB GPU. Provider: DeepSeek. Released 2024-05-15. Access: Open.
Aggregate Rating 1404 ± 17, rank #1189 of 1635 rated models, from 170 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| OpenCompass Language - Content Summarization - Chinese (CompassBench 2405) | 24.1 | Score (%) | 71.1 |
| Open PL LLM - EQ-Bench | 38.98 | EQ-Bench Score | 64.9 |
| Open Portuguese LLM - FaQuAD NLI | 71.28 | Macro F1 (%) | 62.9 |
| OpenCompass Math - Middle School - English (CompassBench 2405) | 18.4 | Score (%) | 60.5 |
| Open PL LLM - PPC (multiple choice, 5-shot) | 71.8 | Accuracy (%) | 57.1 |
| Open PL LLM - CBD (multiple choice, 5-shot) | 23.97 | Macro F1 (%) | 54.2 |
| OpenCompass Subjective - Language - Chinese (CompassBench 2405) | 38.5 | Score (%) | 52.6 |
| OpenCompass Subjective - Language (CompassBench 2405) | 38.3 | Score (%) | 50 |
| Open PL LLM - PPC (generative, 5-shot) | 65.9 | Accuracy (%) | 48.3 |
| CSimpleQA | 36 | Score (%) | 47.5 |
| OpenCompass Agent - T-Eval - Chinese (CompassBench 2405) | 71.2 | Score (%) | 47.4 |
| OpenCompass Subjective - Language - English (CompassBench 2405) | 38.1 | Score (%) | 47.4 |
Interactive version: theaggregate.ai/model?slug=deepseek-v2-lite-chat · How It Works · Data refreshed daily, snapshot 2026-11-05.