AI Potluck
Contribute
Back to Gap Map Model components / Fine-tuned / chat models

StarCoder2

BigCode Project
open weights / Overall score: 2.4

BigCode's self-aligned instruct release of StarCoder2 at 16B, trained with a fully transparent pipeline and evaluated on EvalPlus, LiveCodeBench and DS-1000. The base StarCoder2-15B it builds on was trained on The Stack v2 and belongs to the base_pretrained shelf.

Scored on the instruct model; the declared artifact is the base repository, so its downloads are read as the instruct model's.

Openness

4 high confidence
4.0
weights
open
data
open(self-oss-instruct-sc2-exec-filter-50k + 7 documented intermediate datasets)
code
open(starcoder2-self-align pipeline on GitHub)
base-data
open(The Stack v2)
license
BigCode-OpenRAIL-M(open but RAIL behavioral-use restrictions, not OSI)

StarCoder2 publishes its weights, its full training corpus (The Stack v2) and the self-alignment pipeline behind the instruct checkpoint. Its BigCode-OpenRAIL-M license restricts what the model can be used for rather than who can use it commercially, and because that license is not OSI-approved, it stays at open weights rather than a fully open release.

  • https://huggingface.co/api/models/bigcode/starcoder2-15b-instruct-v0.1 recorded 2026-09-29

    Hugging Face API record for the instruct model shows "gated":false, the tag "license:bigcode-openrail-m", and the tag "dataset:bigcode/self-oss-instruct-sc2-exec-filter-50k".

  • https://huggingface.co/bigcode/starcoder2-15b recorded 2026-09-29

    Base model card: "StarCoder2-15B model is a 15B parameter model trained on 600+ programming languages from" The Stack v2, and "The model is licensed under the BigCode OpenRAIL-M v1 license agreement."

  • https://huggingface.co/bigcode/starcoder2-15b-instruct-v0.1 recorded 2026-09-29

    Model card calls it the "very first entirely self-aligned code Large Language Model (LLM) trained with a fully permissive and transparent pipeline" and says "Our open-source pipeline uses StarCoder2-15B to generate thousands of instruction-response pairs"; it links Code bigcode-project/starcoder2-self-align and Dataset bigcode/self-oss-instruct-sc2-exec-filter-50k, and says "We provide intermediate datasets for every step of the pipeline".

Adoption

1 medium confidence
1.0

Adoption is measured as Hugging Face downloads for the instruct SKU and the base StarCoder2-15B, and both see minimal download volume. The self-alignment recipe is significant to researchers, but that has not translated into wide production use.

Capability

3 high confidence
3.0

The model card reports HumanEval pass@1 of 72.6 and HumanEval+ of 63.4 and calls it the first entirely self-aligned code model trained with a fully permissive and transparent pipeline. It reports no agentic coding results, and it sits below current frontier coding models.

Verified 2026-09-27