StarCoder 2 15B
BigCode
BigCode's open-source code generation model with 15B parameters trained on 4T+ tokens from The Stack v2 dataset covering 600+ programming languages. Features 16,384-token context window enabling handling of longer codebases and elaborate instructions. Trained on 1T tokens of permissively licensed GitHub data including code, Git commits, issues, and Jupyter Notebooks. Achieves 33.6%-44.2% on HumanEval and excels on MultiPL-E across 16 of 18 programming languages. Best-performing large model on DS-1000 benchmark. Outperforms CodeLlama-34B on math and code reasoning while matching low-resource language performance. Released under permissive BigCode OpenRAIL-M license. Collaboration between ServiceNow, Hugging Face, and NVIDIA.
Strengths
- Supports 600+ programming languages - exceptional language breadth
- Best-performing large model on HumanEval, MBPP, and EvalPlus datasets
- Excels on 16 of 18 languages in MultiPL-E benchmark
- 16K context window handles larger codebases than predecessors
- Trained on 4T+ tokens from The Stack v2 - high-quality permissive data
- Outperforms CodeLlama-34B on math and code reasoning
- Permissive OpenRAIL-M license enables broad commercial use
Caveats
- 44% HumanEval lower than GPT-4o (90.2%) and Claude Opus 4.5
- 15B size requires significant GPU memory
- Training data through early 2024 - missing latest language features
- No multimodal capabilities for UI/design understanding
- Consider GPT-4o or Claude for highest-accuracy code generation
- Vendor
- BigCode
- Context Window
- 16,384 tokens
- Max Output
- 4,096 tokens
- License
- BigCode OpenRAIL-M
- Release Date
- 2024-02-01
- Modalities
- text
Capabilities
Pricing
Open source (OpenRAIL-M) - free. Requires GPU. API providers typically $0.25-$0.50 per 1M input tokens