DeepSeek Coder V2 Lite Instruct: benchmark results
DeepSeek's 16B MoE coder (2.4B active) with a 128K context, supporting 338 programming languages. Provider: DeepSeek. Released 2024-06-17. Access: Open.
Aggregate Rating 1453 ± 21, rank #3656 of 10916 rated models, from 39 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| GSM8K | 87.6 | Accuracy (%) | 90.3 |
| Delulu - Edit Similarity | 71.7 | Edit similarity (0-100): character-level normalized Levensht | 90 |
| Sphinx - C# | 38.57 | Checklist Coverage Score (%; GPT-4o judge, 450 buggy + 50 bu | 89.3 |
| Sphinx - Java | 27.76 | Checklist Coverage Score (%; GPT-4o judge, 450 buggy + 50 bu | 89.3 |
| Sphinx - JavaScript | 22.31 | Checklist Coverage Score (%; GPT-4o judge, 450 buggy + 50 bu | 89.3 |
| Delulu - CodeBLEU | 48.9 | CodeBLEU (0-100): n-gram, syntax-tree and dataflow agreement | 80 |
| Delulu - Exact Match | 46.9 | Exact match (%) of the completion with the gold completion, | 80 |
| MERA Code - YABLoCo | 7.21 | pass@1 (%) | 58.5 |
| FullStackBench (2024-12 bilingual) | 48.73 | Pass@1 (%), 3,374 problems: 1,687 English and 1,687 Chinese; | 53.8 |
| Sphinx | 28.09 | Checklist Coverage Score (%; GPT-4o judge, 450 buggy + 50 bu | 50 |
| RedCode - Exec Bash (ReAct) | 68.42 | Attack success rate (%) | 45.5 |
| MERA Code - RealCodeJava | 25.5 | pass@1 (%) | 43.8 |
Interactive version: theaggregate.ai/model?slug=deepseek-coder-v2-lite-instruct · How It Works · Data refreshed daily, snapshot 2026-11-08.