DeepSeek-V2-Lite: benchmark results
DeepSeek's V2-Lite, a 16B-parameter Mixture-of-Experts base model with 2.4B active parameters, trained from scratch on 5.7 trillion tokens using Multi-head Latent Attention and DeepSeekMoE architecture. Provider: DeepSeek. Released 2024-05-15. Access: Open.
Aggregate Rating 1349 ± 33, rank #1360 of 1635 rated models, from 66 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| MERA Industrial - ruTXTAgroBench - Crop cultivation technologies | 69.4 | Score (%) | 96.2 |
| MemPoison - Poisoned Task | 56.96 | Behavioral Corruption Rate (%) | 88.9 |
| MERA Industrial - ruTXTAquaBench - Mariculture | 42.8 | Score (%) | 80.5 |
| MERA Industrial - ruTXTAgroBench - Crop production | 42 | Score (%) | 78.6 |
| EvoHarmBench - Spam and Flooding | 72.3 | ASR@Readable (%; mean over the spam and flooding sub-cluster | 77.8 |
| MERA Industrial - ruTXTAgroBench - Farming systems on agricultural landscapes | 38.1 | Score (%) | 66.7 |
| MERA Industrial - ruTXTAgroBench - Forage production and meadow farming | 31.1 | Score (%) | 63.8 |
| MERA Industrial - ruTXTAgroBench - Reclamation farming | 35.2 | Score (%) | 61.9 |
| MERA Industrial - ruTXTAquaBench - Ichthyopathology | 30.1 | Score (%) | 61.4 |
| MERA Industrial - ruTXTAquaBench - Feeding of fish and aquatic organisms | 29.7 | Score (%) | 47.6 |
| Open PL LLM - PolQA Open Book (generative, 5-shot) | 79.05 | Levenshtein Similarity (%) | 47.3 |
| Open PL LLM - PolQA Closed Book (generative, 5-shot) | 49.53 | Levenshtein Similarity (%) | 44.5 |
Interactive version: theaggregate.ai/model?slug=deepseek-v2-lite · How It Works · Data refreshed daily, snapshot 2026-11-05.