StarCoder2-3B
Release in the StarCoder2 family · version starcoder2-3b
Maintained by BigCode project (open scientific collaboration)16, ServiceNow (BigCode co-steward; trained this model)67, Hugging Face (BigCode co-steward)67
The smallest StarCoder2 base model, with about 3B parameters, trained on more than 3 trillion tokens from The Stack v2 and other sources. Its model card lists 17 programming languages, grouped-query attention, a 16,384-token context window with 4,096-token sliding-window attention, and a fill-in-the-middle objective. ServiceNow trained it for the BigCode project.172
- Model hub: Model card (Hugging Face) (external site: huggingface.co)
- Repository: StarCoder2 repository (external site: github.com)
- Paper: StarCoder 2 and The Stack v2: The Next Generation (arXiv 2402.19173) (external site: arxiv.org)
- License: BigCode OpenRAIL-M v1 License Agreement (external site: huggingface.co)
Availability and license
Overall availability
Downloadable from Hugging Face without gating. Use is subject to the BigCode OpenRAIL-M v1 license, including its use restrictions.123
Availability is separate from permission: read the license before using or redistributing.
BigCode OpenRAIL-M v1 License Agreement (external site: huggingface.co)13
Apache License 2.0 (StarCoder2 repository) (external site: raw.githubusercontent.com)5
BigCode OpenRAIL-M v1 is a custom responsible-AI license. It grants royalty-free rights to use, modify, and share the model, including commercially, but makes compliance with the use restrictions in its Attachment A a condition of the grant and requires those restrictions to be carried into any shared copies or modifications. It does not cover source code or training data. The card adds that the model can reproduce training code verbatim, which may carry its own license and attribution requirements.31
Model-disclosure tier
The model parameters for this release can be downloaded by the public. License terms may still restrict use, redistribution, or commercial use.
The weights are under a license that is not on the rubric's OSI-approved list. Read its terms before use.
Public materials checklist
| Item | Status | Notes and evidence |
|---|---|---|
| WeightsCan the general public download the model parameters for this release? | Public | 12 |
| Inference codeIs code for running the model published? | Public | The card and repository document inference with Hugging Face Transformers and serving with text-generation-inference.14 |
| Training codeIs the code used to train the model published? | Unknown | The StarCoder2 repository publishes only a LoRA fine-tuning script, which the rubric does not count as training code for the base model. The card does not name the pretraining framework, and this catalog did not find the pretraining code for this model published.14 |
| Training-data informationPublic = the training data itself can be obtained. Partial = composition or sources are documented without full access. | Partial | The Stack v2 is on Hugging Face behind a terms-of-use acceptance, with provenance information for each file, but its terms say bulk download of the file contents requires an agreement with Software Heritage and Inria.86 |
| Training recipeAre the training configuration and procedure documented in enough detail to follow? | Partial | The card documents architecture, objective, token count (3+ trillion), 1.2 million pretraining steps, bfloat16 precision, and 160 A100 GPUs; the report describes base training at a 4k context window followed by 16k long-context training.16 |
| Evaluation materialsPublic = evaluation code or prompts that let others re-run the evaluations are published. Partial = results only. | Public | The card and report give benchmark results. The StarCoder2 repository directs users to the public BigCode Evaluation Harness to evaluate StarCoder2 models, and the report documents its prompts and says evaluation code is available through that harness. This catalog did not confirm that the harness covers every reported evaluation.146 |
What it is useful for
Code completion. The card notes it is not an instruction model, so commands such as "write a function that..." do not work well.1
Run and use notes
- The model card reports memory footprints measured with Transformers' get_memory_footprint (model weights only): 12,624.81 MB in full precision, 6,312.41 MB in bfloat16, 3,434.07 MB with 8-bit bitsandbytes quantization, and 1,994.90 MB with 4-bit quantization.1
Organization context
Provenance and derivatives
Trained by ServiceNow for the BigCode project on The Stack v2 (built from the Software Heritage archive, with opt-out requests excluded) and additional sources such as arXiv and Wikipedia.178
- Derived from: The Stack v2 (train split) (external site: huggingface.co) — Pretraining data; no separate catalog record.
Other releases in the StarCoder2 family
- StarCoder2-15BModel-disclosure tier (USASI rubric v0.1): Open-weight
U.S. eligibility
Eligible · basis: U.S.-governed project
Published in the BigCode organization on Hugging Face and trained by ServiceNow, according to Hugging Face's announcement. The StarCoder2 report says BigCode is stewarded by ServiceNow (Santa Clara, California, per its 2025 Form 10-K) and Hugging Face (whose privacy policy names Hugging Face, Inc. as its service provider and says the company is located in the United States). Eligibility rests on these documented organizations, not on individual contributors.167910
Sources
This listing is not an endorsement, a safety assessment, or a federal approval.