starcoder
bigcode/starcoder
Open Source · chat · open-weights
Open
Alert me on changes
Context
8.2K
Max output
—
Weights
Open
API $/1M
—
Modalities
text
Released
24 Apr 2023
License: bigcode-openrail-m · bigcode/starcoder
AI summary
● machine-written
StarCoder: Open-weight code generation model with 8K context window
StarCoder is a 15.5B-parameter open-access language model for code generation created by the BigCode community, trained on permissively licensed GitHub code from The Stack dataset. The model supports fill-in-the-middle editing and handles inputs up to 8,192 tokens, shipping under the BigCode OpenRAIL-M license which permits commercial use under responsible AI conditions. StarCoder2, released in 2024, offers three sizes (3B, 7B, 15B) with extended capabilities including 16K context window and support for 600+ languages.
What's new
- Fill-in-the-middle editing for inline code completion and refactoring
- Trained on permissively licensed code from The Stack dataset
- Supports multiple programming languages across HumanEval, MBPP, and DS-1000 benchmarks
- Available for deployment via Transformers, GGUF/Ollama, Hugging Face Inference Endpoints, and NVIDIA NGC containers
- Transparent training pipeline with published preprocessing scripts and opt-out policies
Best for
Code completion and inline editing within development workflowsMulti-language code generation and cross-language translation tasksLocal deployment and VPC-isolated inference scenariosEnterprise adoption with commercial-friendly open-source licensing