StarCoder2
The open-science code model family trained on The Stack v2's 619 languages.
About
StarCoder2 (February 2024) is BigCode's open-science code model family — 3B from ServiceNow, 7B from Hugging Face, 15B from NVIDIA — trained on 3.3–4.3T tokens from The Stack v2, with full training-data transparency via Software Heritage IDs. The 15B model matched or beat CodeLlama-34B at twice its size on the paper's benchmark suite. Released under the BigCode OpenRAIL-M license. Later Qwen and DeepSeek coders beat it on raw benchmarks, but it remains the reference for fully transparent, responsibly-sourced code models.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Team-reported (paper): StarCoder2-15B matches or outperforms CodeLlama-34B; StarCoder2-3B beats other code models of similar size and StarCoderBase-15B.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 7/10
Scored 7/10 (verified tier, metrics-v1). Evidence: The Decoder + BigCode paper (arXiv 2402.19173) read this session: Feb 28 2024, 3B/7B/15B (ServiceNow/HF/NVIDIA), The Stack v2 (619 languages), 3.3-4.3T tokens, BigCode OpenRAIL-M.; Team-reported: 15B matches/beats CodeLlama-34B; 3B beats StarCoderBase-15B..