Qwen2.5-Coder-3B
Alibaba · 2024-11-12 · 3.1B parameters
Qwen2.5-Coder-3B is Alibaba's 3,085,938,688-parameter base code language model, released with the expanded Qwen2.5-Coder family on November 12, 2024. Qwen describes a code-specialized continued-pretraining pipeline using about 5.2 trillion file-level tokens followed by approximately 300 billion repo-level tokens; the public-repository source-code component spans 92 programming languages. This text-only checkpoint supports a 32,768-token context window and is released under the Qwen Research license.
Benchmark scores
| Benchmark | Score |
|---|---|
| ARC AI2 | 52.9 |
| Capability | 102.8 |
| Epoch Capabilities Index (ECI) | 107 |
| GSM8K | 75.7 |
| HellaSwag (Unspecified Scoring Protocol) | 70.9 |
| MMLU | 61.2 |
| WinoGrande | 67.4 |