PolyCoder
CMU's 2022 code LLM — the first fully open Codex-era code model
Overview
PolyCoder is a 2.7B GPT-2-style code model trained from scratch at Carnegie Mellon on 249GB of code in 12 languages, released in 2022 with weights, data recipe and training code fully open. Best known for beating Codex at C generation despite being a fraction of the size.
Pricing
Pricing shown for reference only. These figures reflect RECATOOLS research as of 11 Jul 2026 and may be out of date or incomplete.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
ASEAN Perspective
PolyCoder in Southeast Asia
ASEAN-region availability and pricing notes coming soon. Drop the editorial team a note via /contact/ if you can supply local context (Singapore/Malaysia/Indonesia/Thailand/Vietnam).
PolyCoder matters for what it opened, not what it can do today. The 2022 Carnegie Mellon model (160M to 2.7B parameters, GPT-2 architecture, 249GB of code in 12 languages) was the first code LLM to ship weights, data pipeline and training code together, back when Codex was a sealed API — and it famously beat Codex at C generation. That makes it a solid, citable baseline for code-model research and teaching. It is not a coding assistant: no product, no API, no IDE integration, no support, and capability far below Code Llama, DeepSeek-Coder or any modern copilot. Self-hosting is the only option. Recommended for researchers; irrelevant for working developers.
What people say
The result people still cite: a 2.7B-parameter model trained by CMU researchers beat OpenAI's Codex at generating C code. That was the headline from "A Systematic Evaluation of Large Language Models of Code" (Xu, Alon, Neubig and Hellendoorn, February 2022), and PolyCoder was the artifact — a GPT-2-architecture model trained from scratch on 249GB of GitHub code across 12 languages, released in 160M, 405M and 2.7B sizes.
Its real contribution wasn't the C benchmark, it was openness. At the time Codex was API-only and its training data a black box; PolyCoder shipped weights, the data pipeline and training code in the VHellendoorn/Code-LMs repo, which is why Slashdot billed it as the first open-source code-generating AI model. Anyone could inspect what went in and reproduce what came out — a norm the field only later caught up to.
Four-plus years on, treat it strictly as a research baseline. The repo is a paper companion, not a maintained project; there's no API, no IDE plugin, and capability sits far below Code Llama, StarCoder, DeepSeek-Coder or Qwen-Coder, let alone commercial assistants. Students and researchers who want a small, fully inspectable code LLM to probe or fine-tune still have a genuine use for it. Everyone else is looking at a museum piece — an important one.
Summary of public user & expert reviews, compiled by RECATOOLS.
Notable facts
- PolyCoder outperformed GPT-3 on C code generation despite being 12x smaller, demonstrating that domain-specialised training beats scale for narrow tasks.
- The model was trained on a personal compute budget at CMU — demonstrating that a small academic team could build competitive code models without Big Tech resources.
- PolyCoder was one of the first papers to systematically document how code model quality varies across programming languages based on training data composition.
Frequently asked questions
About this listing
This entry was compiled from publicly available data including PolyCoder's official website, press releases, documentation, and reputable third-party publications. RECATOOLS is not affiliated with PolyCoder 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 PolyCoder directly →
Spotted something out of date? Suggest an update →
PolyCoder in the news
Alternatives to PolyCoder
More in Code & Dev Tools
