StarCoder
BigCode's self-hostable code models with traceable training data
Overview
StarCoder is the code-generation model family of BigCode, a collaboration led by Hugging Face and ServiceNow. StarCoder (15B, 80+ languages) shipped in May 2023 and StarCoder2 (3B, 7B and 15B, trained on The Stack v2) in February 2024. The weights are released under the BigCode OpenRAIL-M licence, which permits commercial use with use restrictions.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete. This is not financial or purchasing advice — always confirm the current price on the provider’s official website before making any decision.
Use cases
What you can produce with StarCoder
- StarCoder2 base models at 3B, 7B and 15B parameters with a 16,384-token context
- Fill-in-the-middle code completion across 600+ languages (15B) or 17 languages (3B and 7B)
- A self-aligned StarCoder2-15B-Instruct model for instruction-style coding prompts
- The original StarCoder and StarCoderBase 15B models from 2023
- A search index for tracing generated code back to training data and its licence
- Fine-tuning scripts for adapting the models to a private codebase
StarCoder2 is a capable mid-size code model whose real strength is provenance: training data filtered to permissive and unlicensed code, a public search index for attribution, and a licence that spells out what is and is not allowed. The 15B model matched CodeLlama-34B at release.
It has not moved since 2024, newer open coders such as Qwen2.5-Coder score well above it, and no hosted provider serves it, so you supply the GPUs. Choose it when documented data governance matters more than benchmark position.
What we found
StarCoder is the code-model family of BigCode, an open scientific collaboration led jointly by Hugging Face and ServiceNow. The first generation arrived on 4 May 2023: StarCoderBase, a roughly 15-billion-parameter model trained on one trillion tokens from more than 80 programming languages, Git commits, GitHub issues and Jupyter notebooks, and StarCoder, the same model further trained on 35 billion Python tokens. Its context of more than 8,000 tokens was, by BigCode's account, longer than any other open model's at the time.
StarCoder2 followed on 28 February 2024 in three sizes, each trained by a different partner: 3B by ServiceNow, 7B by Hugging Face and 15B by NVIDIA. The 15B model saw more than 4 trillion tokens across 600-plus languages from The Stack v2, built with Software Heritage; the two smaller models cover 17 languages. All three have a 16,384-token context with 4,096-token sliding-window attention. The base models are not instruction-tuned, though BigCode added a self-aligned StarCoder2-15B-Instruct in April 2024. The StarCoder2 paper reports the 15B model matching or beating CodeLlama-34B, more than twice its size, but newer models have moved past it: the Qwen2.5-Coder technical report lists StarCoder2-15B at 46.3 on HumanEval against 61.6 for Qwen2.5-Coder-7B.
The licence is BigCode OpenRAIL-M, which allows commercial use with use restrictions attached, among them no malware, no medical advice that substitutes for professional care, and disclosure when machine-generated content is published. BigCode itself says it is not an open-source licence by the OSI definition. Training data was filtered to permissively licensed and unlicensed code, and BigCode publishes a search index for tracing generated code back to its source.
Development has stopped. The StarCoder2 GitHub repository was last pushed in March 2024, no inference provider on Hugging Face serves the models, and NVIDIA deprecated its hosted StarCoder2-15B endpoint on 24 July 2025. What remains is a self-hostable code model with unusually transparent training data.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- StarCoder was trained on The Stack dataset, which was filtered using an opt-out mechanism — developers could request their GitHub repositories be excluded from training.
- The 15B parameter model achieves performance comparable to closed 30B code models, demonstrating that careful data curation matters more than raw scale.
- BigCode is a unique collaboration between two companies (Hugging Face and ServiceNow) and the broader academic community, releasing all training data and model weights openly.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including StarCoder's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with StarCoder unless explicitly stated.
Third-party AI tools update their pricing, features, availability, and policies frequently. Information here may be outdated by the time you read this — we make reasonable efforts to keep listings current, but cannot guarantee absolute accuracy.
For the latest details, please refer to StarCoder directly →
Spotted something out of date? Suggest an update →
Alternatives to StarCoder
More in Code & Dev Tools
