starcoder2-3B: benchmark results
BigCode's StarCoder2-3B, a 3B code model trained on 17 programming languages from The Stack v2 using Fill-in-the-Middle over 3+ trillion tokens, not an instruction-following model. Provider: BigCode. Released 2024-02-28. Access: Open.
Aggregate Rating 1189 ± 20, rank #1596 of 1656 rated models, from 34 benchmark results.
Strongest benchmark results
| Benchmark | Score | Metric | Percentile |
|---|---|---|---|
| EffiBench - NET | 2.9 | Normalized Execution Time | 70.7 |
| Open LLM Leaderboard v1 - GSM8K | 19.64 | Accuracy (%) (5-shot) | 40.7 |
| BigCode Models Leaderboard | 31.4 | HumanEval Python Pass@1 (%) | 25.4 |
| ShaderMatch | 7.36 | Clone Match Rate (%) | 23.8 |
| GSM8K | 21.6 | Accuracy (%) | 23.7 |
| CRUXEval | 34.2 | Output Prediction pass@1 (%) | 21.4 |
| Open LLM Leaderboard v1 - MMLU | 38.65 | Accuracy (%) (5-shot) | 20.3 |
| Open LLM Leaderboard - BBH | 35.09 | Score | 19.3 |
| EvoEval Concise | 33.54 | Pass@1 (%) | 18 |
| EffiBench - NMU | 1.94 | Normalized Memory Usage | 17.1 |
| Open LLM Leaderboard - MMLU-Pro | 16.36 | Score | 17.1 |
| EvoEval Creative | 14 | Pass@1 (%) | 17 |
Interactive version: theaggregate.ai/model?slug=starcoder2-3b · How It Works · Data refreshed daily, snapshot 2026-10-20.