Frontier models and agents.Onchain systems and markets.
Qomira Labs is a research company with two programmes that stand on their own. Models for video, audio, image, and text, and agents. Onchain protocols, markets, blockchain infrastructure, and quantitative research. We do the core work once, and our products take it to the people who use it.
The hard part of this work is expensive and it does not get cheaper by being done repeatedly. So we do it in one place.
Qomira Labs holds the research, the models, and the compute. Each product built on top of it goes deep on a single problem and ships to real users. Research does not have to justify itself quarter by quarter, and products do not have to rebuild foundations they should be able to assume.
The meter is inside the seller's building, and it has no window.
The systems are already shipping
Open-weight models went from roughly 30 percent of routed tokens to about 61 percent in under a year, and money moves onchain in volume every day. What anyone can run, anyone can deploy, to any user, in any language, without asking permission.
The party being paid holds the meter
You are told how many tokens you used and cannot recount them. Market size is read off chain data that is mostly not what it appears to be. Safety is evaluated in English by the people who built the system. In each case the number that matters is produced by whoever benefits from it.
We build the measurement first
The evaluation before the model, the benchmark before the claim, published with the results our own systems fail. Open weights, because a model nobody is allowed to run has not reached anybody.
The swing in benchmark score from serving the same open weights through a different backend. Buyers cannot tell which one answered them.
arXiv 2605.19537The share of onchain stablecoin movement that is an actual payment. Anyone sizing this market from chain data is wrong by two orders of magnitude.
McKinsey and Artemis, via published analysisBilling inflation achieved against a published audit scheme without being detected. Every audit in that class checks something the seller produced.
arXiv 2605.30040Qomira Labs builds the technology. Each product below owns one problem and ships.
The internet capital market for on-chain assets.
Trade, borrow, and access tokenized real-world assets on one unified liquidity layer. Liquidity that never sleeps.
VISIT JUNCTAFull velocity for every token.
VELORAK makes AI inference faster and cheaper on any GPU. Point your app at one endpoint, and VELORAK routes, caches, tunes and bills every token, so you pay less and your users wait less.
We publish when the work is ready. There is nothing here yet.
Subscribe for regular updates on what we publish and release.
What will land here: benchmarks and the evaluation sets behind them, model and dataset releases with the weights, measurements of what things actually cost to run, and the results our own systems fail. Negative results included, because a lab that publishes only its successes teaches everyone to hide.
Qomira Labs is a research company working across artificial intelligence and onchain systems.
Frontier capability is arriving faster than the measurement that would tell anyone whether it works. The gap is widest exactly where the users are least able to absorb a failure: in languages nobody evaluates, on devices nobody optimises for, over rails whose volume is counted by the party being paid.
We do the expensive research once and let each product draw on it. We build the evaluation before the model and publish it, including the results our own systems fail. We release weights, because a model nobody is allowed to run has not reached anybody.
Five programmes hold the work. Frontier models and agents across text, audio, image and video. Onchain protocols, markets and quantitative research. The infrastructure, data and cost work underneath both. Safety, privacy and governance, which is the fastest-moving of them, because capability is arriving faster than anyone's ability to say whether it is safe for the people receiving it. And two standing questions we have deliberately not answered yet: whether to design hardware, and how far to go into physical AI.
Research here does not have to justify itself quarter by quarter. Every agenda we hold ends with a section on what would make it wrong, and those sections are the point rather than a disclaimer.
We work with labs, institutions and builders of any size.