Numerical tools that work across scales.
Multilevel Labs is an early-stage research company. We apply multilevel methods to hard problems in science and engineering, and we turn the results into tools that the people doing the work can actually use.
Coarse structure and fine detail, solved together.
Most hard problems in science and engineering live on many scales at once: the slow, collective behaviour of a system and the fast, local details that drive it. Methods that only see one scale spend most of their effort fighting the other.
Multilevel methods treat the scales as one problem. We apply that idea, refined over decades in numerical analysis, to the computational tools that researchers rely on every day, and we deliver the results as libraries and APIs rather than papers alone.
Current focus is on neural network optimization.
How we operate.
For the whole industry.
We're building an algorithmic platform, not a feature for one company. No customer or hardware partner gets long-term exclusivity, so efficiency gains spread across the industry instead of stopping at one company.
Careful with what we release.
Cheaper computation changes what people can build with it. We take that into account in what we release, who we work with, and the terms we agree to.
Numbers you can reproduce.
Any performance figure we give can be reproduced, under the conditions we state, by whoever we gave it to.
On our customers' side.
We don't use a customer's data or insight against them, and we don't build products that compete in their markets.
A long horizon.
We are a Public Benefit Corporation. Commercial work funds long-term research, and we measure our mission over ten years by the computing cost our technology removes. We will report on it publicly each year.
Talk to us.
We're a small team and mostly heads-down. If you have a specific reason to get in touch, this is the place. We read everything and reply when we can.