Qoru is live: building open, regional-first language models
Today we're announcing the official launch of Qoru — a research lab building open-weight language models that work well for underserved languages and communities worldwide.
Today we're pulling back the curtain on a project three years in the making. Qoru is officially live. We're a research lab building open-weight language models that work well for underserved languages and communities worldwide — not just English and Mandarin.
I'm Dan, founder and chief ML engineer. A few years ago I was fine-tuning English-only checkpoints for African languages and watching promising results evaporate the moment a sentence left the training corpus. Every day since, my co-founder Paul and I have been building the tooling, datasets, and yes — models — to close that gap. Today we ship the first results.
The problem we're fixing
The AI ecosystem is heavily skewed toward high-resource languages. A model that handles English well but stumbles on Igbo, Swahili, or Yoruba isn't a general-purpose model — it's a narrow one with global ambitions. Yet the teams building for those languages have no shortage of creativity, just an unfair deficit of compute and data.
We believe that's a solvable engineering problem, not a fundamental limit.
What Qoru builds
Our stack has three legs:
-
Data infrastructure for low-resource languages. We've built ingestion pipelines that pull from community archives, government documents, and parallel corpora that most teams never touch. Cleaning and de-duplicating these sources reliably is non-trivial — and it's the foundation everything else sits on.
-
Efficient fine-tuning at scale. Techniques like LoRA and QLoRA let us specialize large base models with a fraction of the compute. The math is public, but the recipes that actually converge are not. We're sharing ours.
-
Open-weight releases with full transparency. Every model we ship comes with training scripts, evaluation harnesses, and model cards. We release under Apache 2.0 for code and permissive licenses for weights. No signup walls.
What's available today
Our first public release is a family of multilingual instruction-fine-tuned models built on top of community checkpoints, optimized for 11 languages spanning West Africa, East Africa, and Southeast Asia. Weights and code are live on huggingface.co/qoru.
We're also publishing the datasets — curated and cleaned — so other teams can build on our work rather than recreating it. Details are in our companion post on the data pipeline.
What's coming next
We have three things on the roadmap:
- A hosted inference API for our open models, with generous free tiers for researchers and non-commercial projects.
- Community grants — a quarterly program funding individuals and small teams building with our models in underserved language communities.
- The next model family, trained from scratch on a broader multilingual corpus. We're targeting 2027 for the first checkpoint.
Join us
We're a small team — Myself, Paul, and a handful of collaborators — but we're working in the open. If you're building for or researching underserved languages, we want to hear from you. Reach out on Twitter, GitHub, or email hi@qoru.ai.
The models ship today. The work starts now.