Small models. Big access. Open science.
We believe the future of AI shouldn't be locked behind API gates and billion-dollar compute budgets.
Nova Research is an open-source AI lab focused on building small, efficient language models that anyone can run, study, and build on — on their own hardware, on their own terms.
The prevailing assumption is that intelligence scales with size. We disagree — or at least, we think the story is incomplete.
A 3B parameter model trained on the right data, with the right architecture choices, can outperform models 10x its size on tasks that actually matter to real people. The research community has shown this repeatedly (Phi, TinyLlama, OpenHermes), but there's still massive unexplored territory.
We exist to push that frontier: How small can we go and still be genuinely useful?
- Fine-tuned specialist models — open-weight models tuned for real-world tasks where small models can compete with or beat large general-purpose APIs
- Training research — novel data curation strategies, efficient training recipes, and architectural experiments in the 1B–7B parameter range
- Accessible tooling — serving infrastructure, quantization guides, and deployment templates so our models are useful on day one, not just benchmarked
Everything we ship is open-weight, open-data, and documented with full training methodology.
Open by default. Models, data, training configs, and results are public. Science that can't be reproduced isn't science.
Useful over impressive. We optimize for real-world utility, not leaderboard position. A model that helps a developer ship faster or a student learn better matters more than a MMLU score.
Small is a feature. Running locally, on a laptop, on a $5/month VPS — that's not a limitation, it's the whole point. If you need an H100 cluster to use it, we haven't done our job.
Show the work. Every release includes a model card, training details, benchmark results, and a write-up on what we learned — including what didn't work.
| Phase | Focus | Status |
|---|---|---|
| Phase 1 | Fine-tune existing open models (Llama, Qwen, Phi, Gemma) for underserved use cases using QLoRA/LoRA | 🔜 Next |
| Phase 2 | Publish training recipes, data curation pipelines, and reproducible benchmarks | Planned |
| Phase 3 | Train custom small models from scratch with novel data strategies | Planned |
| Phase 4 | Explore architectural innovations for efficient inference on consumer hardware | Research |
- Training: Hugging Face Transformers, TRL, Axolotl, PyTorch
- Serving: vLLM, llama.cpp, FastAPI
- Tracking: Weights & Biases
- Distribution: Hugging Face Hub (models, datasets, spaces)
We're early. If you believe AI should be accessible, auditable, and owned by the people who use it — not just the companies that build it — we'd love to hear from you.
- ⭐ Star this repo to follow along
- 🐛 Open an issue with ideas, feedback, or collaboration proposals
- 📝 Read our write-ups on [coming soon]
- 🤝 Reach out if you want to contribute research, compute, or code
Our models and code are released under permissive open-source licenses (Apache 2.0 / MIT) unless otherwise noted per release.
"A nova is a star that suddenly becomes thousands of times brighter — then keeps shining."