PyCodeGPT
Python-specialised code generation model — trained exclusively on high-quality Python to maximise Python coding quality.
Overview
PyCodeGPT is a 110M-parameter GPT-Neo-based Python code model from Microsoft Research, trained on 13M files (96GB) filtered from 1.2M GitHub repos and released alongside the 2022 CERT paper. A research baseline with open weights, not a maintained product.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 1 Sep 2026 and may be out of date or incomplete.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 PyCodeGPT
- A 110M-parameter Python-specialised model small enough to fine-tune on a single GPU
- A documented data-filtering recipe: 96GB of Python distilled from 1.2 million GitHub repositories, with a fresh 32K-vocabulary tokenizer
- A clean baseline for Python-specialisation experiments
- The CERT work it shipped as the vehicle for — an IJCAI 2022 paper on continual pre-training for library-oriented generation, aimed at pandas- and NumPy-specific completion
- ⚠️ 8.32% pass@1 on HumanEval. Respectable for a 110M model in 2022; today free assistants clear ten times that
- ⚠️ Not a coding tool, and not quite where you would expect to find it: the weights sit on a researcher's personal Hugging Face account (Daoguang/PyCodeGPT) rather than Microsoft's org, and the last release dates to October 2022
PyCodeGPT is a 110M-parameter Microsoft Research model from 2022, trained purely on filtered GitHub Python and released alongside the CERT paper on library-oriented code generation. Its interest is academic: a clean, small, fine-tunable baseline for studying language-specialised code models, with documented data curation and open weights — hosted, oddly, on a researcher's personal Hugging Face account rather than Microsoft's. It is not a practical assistant. Its HumanEval pass@1 of 8.32% was creditable in 2022 and is nowhere near modern Qwen, DeepSeek or StarCoder-family coders, and there's no hosted product, pricing or support. Last release: October 2022. Use it as a reference model or teaching artifact, nothing more.
What people say
The numbers tell you what this is: 110 million parameters, 96GB of Python filtered down from 1.2 million GitHub repositories, and 8.32% pass@1 on HumanEval. PyCodeGPT is a small Microsoft Research model from 2022, built on GPT-Neo with a fresh 32K-vocabulary tokenizer, that made a focused bet — train on one language, curated hard, and match Codex at the same size. It roughly did: the repo reports it comparable to similarly sized Codex variants and ahead of CodeParrot.
It shipped as the vehicle for CERT, an IJCAI 2022 paper on continual pre-training for library-oriented code generation — think pandas- and NumPy-specific completion. That paper is why the model exists, and it's where the lasting research interest lives. One quirk worth knowing: the weights sit on a researcher's personal Hugging Face account (Daoguang/PyCodeGPT), not Microsoft's org, and the repo's last release dates to October 2022.
Nobody should mistake this for a coding tool. An 8% pass@1 was respectable for a 110M model in 2022; today free assistants clear ten times that. But as a research artifact it still has legitimate uses — it's small enough to fine-tune on a single GPU, the data-filtering recipe is documented, and it makes a clean baseline for Python-specialisation experiments. Reference model, not a Copilot replacement.
Summary of public user & expert reviews, compiled by RECATOOLS.
Notable facts
- PyCodeGPT was the research precursor to Microsoft's Phi-1 model, which extended the quality-filtering approach to educational text and demonstrated it at a larger scale.
- The model showed that training on 500k high-quality Python files could outperform training on 5 million lower-quality Python files.
- Microsoft released PyCodeGPT as part of a broader research programme to understand how data quality affects code model performance — research that directly informed GitHub Copilot improvements.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including PyCodeGPT's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with PyCodeGPT 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 PyCodeGPT directly →
Spotted something out of date? Suggest an update →
Alternatives to PyCodeGPT
More in Code & Dev Tools
