StarCoder2-3|7|15B logo

StarCoder2-3|7|15B

Free

Open-source code generation models by BigCode

FreeFree tier
Inputs: textOutputs: text
Type
Open Source

About StarCoder2-3|7|15B

StarCoder2 is a family of open-source large language models for code, developed by the BigCode project. Available in three sizes (3B, 7B, 15B parameters), these models are fine-tuned on The Stack v2 dataset and are optimized for code generation, completion, and assistance. The models were released in February 2024 and are freely accessible on Hugging Face under permissive licenses.

Key Features

Available in 3B, 7B, and 15B parameter sizes
Fine-tuned on The Stack v2 dataset for code understanding
Supports code generation, completion, and infilling
Released under open-source license
Trained on permissively licensed code from multiple programming languages

Pros & Cons

Pros
  • Free and open-source with multiple model sizes to suit different resource constraints
  • Strong performance on code tasks compared to similarly sized models
  • Trained on a large, deduplicated code dataset (Stack v2)
  • Backed by the BigCode community and widely supported on Hugging Face
  • Supports multiple programming languages
Cons
  • Requires significant GPU memory for the 15B model (e.g., ~30GB VRAM for inference)
  • Primarily designed for code tasks; limited performance on general NLP tasks
  • May exhibit biases or errors typical of large language models

Best For

Code generation from natural language descriptionsCode completion and autocomplete in IDEsCode explanation and documentation generationProgramming assistance and debuggingEducational tool for learning programming

FAQ

What sizes are available for StarCoder2?
StarCoder2 is available in three sizes: 3B, 7B, and 15B parameters.
Is StarCoder2 free to use?
Yes, StarCoder2 models are open-source and freely available on Hugging Face.
When was StarCoder2 released?
The models were published in February 2024, with a paper (arXiv:2402.19173) released on February 29, 2024.
What dataset was StarCoder2 trained on?
StarCoder2 was trained on The Stack v2, a large dataset of permissively licensed source code.
Who developed StarCoder2?
StarCoder2 was developed by the BigCode project, a collaboration between Hugging Face and ServiceNow.