StarCoder2
Fully transparent open code LLMs in 3B/7B/15B — trained on The Stack v2 with every source traceable.
About
StarCoder2, released February 2024 by the BigCode open-science collaboration (ServiceNow, Hugging Face, NVIDIA), is the next generation of transparently trained code LLMs. Three sizes cover three needs: 3B (trained by ServiceNow on 17 languages, 3T+ tokens), 7B (Hugging Face, 3.5T+ tokens), and 15B (NVIDIA NeMo on 600+ languages, 4T+ tokens) — all using grouped-query attention and fill-in-the-middle training.
Its distinguishing feature is radical transparency: trained on The Stack v2 (built on Software Heritage's source archive), the project releases model weights, datasets, processing code, and training code, with Software Heritage persistent IDs for the source data. The 3B model matches the original 15B StarCoder, and the 15B matches or beats CodeLlama-34B — more than twice its size. Weights use the BigCode OpenRAIL-M v1 license, which carries responsible-use restrictions rather than pure open-source terms.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
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 (metrics tier, metrics-v1). Evidence: Verified via vendor site/docs/GitHub and reputable coverage, wave 59; Trending score 68/100 from researcher assessment.