Train & serve
LoRA adapters have been trained and exported, with model loading, inference, and tool use tested in development.
OUR MODELS
Open models.
Made to specialize.
For people who want trained models they can understand, adapt, and run themselves. Our first focus is Python; our longer-term goal is capable models for low-end hardware.
- Foundation
- Gemma 4 · 31B
- Method
- LoRA fine-tuning
- Focus
- Python development
- Stage
- Research prototype
PROJECT 001 · IN DEVELOPMENTA Python
A Python
specialist.
We’re adapting Gemma 4 to navigate Python repositories, make targeted code changes, and verify repairs with tests.
This is our project for Google’s Gemma 4 Developer Agent Competition on Kaggle. Public model downloads and a release model card are still to come.
View the Kaggle competition01CURRENT PROGRESS
From training
to working repairs.
Our Colab experiments have reached adapter training, model inference, and successful individual Python repairs. These are development checks; broader, reproducible evaluation is still ahead.
Repair & verify
Individual repository fixes have passed tests. We’re investigating reliability across more tasks and the competition submission workflow.
Measure & release
Publish code, model artifacts, evaluation methods, and hardware requirements so others can reproduce and build on the work.
No validated competition score or low-end hardware performance claim is announced. The current competition prototype uses a 31B foundation; smaller hardware is a research direction.
02OUR RELEASE APPROACH
More than
a download.
We intend to release our work openly: source code, model artifacts, and research documentation, with clear licenses and enough context to make them useful.
Clear model cards
Intended uses, training context, known limitations, and practical guidance—including memory and hardware requirements.
Meaningful evaluations
Results connected to specific tasks, with enough context to understand what they measure and what they don’t.
Explicit licensing
New original code will use 0BSD. Existing MIT code and upstream model licenses remain intact, with terms identified for every release.
A FEW THINGS TO KNOW
Open, with
clarity.
What kinds of models will FortunaML train?
We’re starting with a Gemma 4 Python specialist for repository navigation and software repair. Future work will explore specialized models with lower hardware requirements.
Will every model be open source?
Our commitment is to release our models, original code, and research openly. New original code will use the permissive 0BSD license; the existing Colab code uses MIT. Gemma 4 is licensed under Apache 2.0, and derived releases will retain applicable upstream terms and notices. Each release will document its model, code, and data licenses.
Will the Python specialist run on my hardware?
Low-end hardware is a priority, but we haven’t established a supported consumer hardware configuration yet. The current competition prototype uses Gemma 4 31B. A release will include measured memory requirements, supported setups, and inference guidance.
Can we collaborate on a specialized model?
We welcome conversations about interesting domains, research questions, and collaboration. Get in touch and tell us about the problem you have in mind.
GOOD THINGS BEGIN WITH A CONVERSATION
A little curiosity.
A lot of possibility.
Research partnerships, funding,
and a shared belief in open intelligence.