OpenGCM

---

About

OpenGCM (Open Generative Coding Models) is an independent, open-source AI research organization.

Our main focus is Hydrion — a series of language models we pretrain from scratch, exploring what small models can do when you build them yourself instead of fine-tuning someone else's.

We also maintain GCM, a line of code-focused chat models built by fine-tuning existing open-weight models.

Everything we release — weights, training data, and methods — is public.

---

Hydrion

Our pretrained model series, trained from the ground up rather than fine-tuned.

Hydrion v3 is a from-scratch pretrain built on a curated, leaderboard-informed data mix (FineWeb-Edu, DCLM-Baseline, FineMath, Cosmopedia v2, and more), with a progressive curriculum that ramps math and reasoning data over the course of training.

---

GCM

Code-focused chat models, fine-tuned on top of existing open-weight base models.

GCM Mark II scores 74.4% pass@1 on HumanEval and 62.2% pass@1 on MBPP.

---

Vision

Small models don't have to mean small ambitions. We're building on a few principles:

• Open — every weight, dataset, and training method we use is public, not just the final model
• Efficient — Hydrion is trained on powerful GPUs, but can be ran on consumer GPUs
• Customizable — models built to be forked, fine-tuned, and taken apart
• Accessible — no gatekeeping behind closed weights or paywalled APIs

---

Research

We're currently focused on:

• Pretraining from scratch — Hydrion, our from-scratch language model series, and the data/architecture decisions that go into it
• Fine-tuning — GCM, our code-focused chat models built on existing open-weight bases
• Evaluation — benchmarking against public leaderboards and tracking what actually moves the needle
• Local deployment — making sure what we build can run on real, non-datacenter hardware

Follow along as it happens in our Research space.

---

Hardware

Our team has:

* Subject to change

---

Team

Built by MrGuineaBird, solhost, and contributors.

---

Sponsored by