SantaCoder
BigCode's small 2022 code model, a data-filtering experiment superseded by StarCoder
Overview
SantaCoder is a 1.1-billion-parameter code model released by the BigCode collaboration in December 2022, trained on the Python, Java and JavaScript subset of The Stack with fill-in-the-middle infilling. It was built to test data-filtering choices, and its findings shaped BigCode's StarCoder models, which have since replaced it. The weights remain free on Hugging Face under the BigCode OpenRAIL-M licence.
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 These figures reflect RECATOOLS research as of 14 Oct 2026 and may be out of date or incomplete.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 SantaCoder
- A 1.1B-parameter code model that runs on modest hardware
- Fill-in-the-middle completion for Python, Java and JavaScript
- Ablation checkpoints on separate branches for comparing filtering and architecture choices
- A paper documenting how deduplication and star-based filtering affect code models
- A search index for tracing generated code back to its training source for attribution
SantaCoder is best read as an experiment that worked: a small model showing that careful deduplication beats popularity filtering, and outscoring larger 2022 code models on MultiPL-E. Its fill-in-the-middle training became standard in BigCode's later work.
For actual coding help it is obsolete, with three languages and a 2,048-token context against StarCoder2's hundreds of languages and 16,384 tokens. The weights and ablation checkpoints remain free, which makes it useful for teaching and research on small code models.
What we found
SantaCoder is a 1.1-billion-parameter code model released in December 2022 by BigCode, an open scientific collaboration supported by Hugging Face and ServiceNow. It was trained on the Python, Java and JavaScript portions of The Stack v1.1, a corpus of permissively licensed code from which owners could opt out, using multi-query attention, a 2,048-token context and the fill-in-the-middle objective, so it can complete a gap between existing code rather than only continue from the end of a file.
It was primarily an experiment. The accompanying paper, posted in January 2023 under the title "SantaCoder: don't reach for the stars!", reports that more aggressive near-duplicate filtering improved results, and that keeping only files from repositories with five or more GitHub stars made them significantly worse. Its best model beat the larger InCoder-6.7B and CodeGen-Multi-2.7B on the MultiPL-E benchmark in both left-to-right generation and infilling across its three languages. Those findings fed directly into BigCode's later StarCoder models.
Those successors are the reason not to use SantaCoder for real work. StarCoder2 comes in 3B, 7B and 15B sizes, with the 3B trained on 17 programming languages and the 15B on more than 600, a 16,384-token context and three to four trillion training tokens.
SantaCoder remains easy to obtain. The ungated weights are still on Hugging Face, with the experimental checkpoints on separate branches and the default checkpoint trained for 236 billion tokens, twice as long as the ablations. The repository was last changed in October 2023, and the hosted demo Space currently fails to start. It is a base model rather than an instruction model, needs trust_remote_code, and writes its fill-in-the-middle tokens with hyphens where StarCoder uses underscores. The BigCode OpenRAIL-M licence allows commercial use with use restrictions and is not OSI-approved, so check its use restrictions before redistribution.
Compiled by RECATOOLS from public sources: vendor documentation and pricing, code repositories and published reporting.
Notable facts
- SantaCoder was released in December 2022 and named after Santa Claus as a holiday release — a whimsical naming convention that caught on in the open-source community.
- At 1.1B parameters, SantaCoder is small enough to run inference on a laptop, making it accessible for developers without dedicated GPU hardware.
- The model was trained on 236 billion tokens of Python, Java, and JavaScript code — approximately 3 million GitHub repositories filtered for quality.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including SantaCoder's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with SantaCoder 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 SantaCoder directly →
Spotted something out of date? Suggest an update →
Related reading
Alternatives to SantaCoder
More in Code & Dev Tools
