Numaro / Autonomous research

Hard problems
need new tools.

We build AI that explores ideas, creates tools, and checks its own work. From mathematical discovery to faster AI and computation.

Explore AI Performance See the research
Explore. Build. Verify. Remember.
From the labA 22-year mathematical record, broken.

AI Performance

Make your hardware go further.

Explore the work
01
Separate operations → one kernel

AI workloads & kernels

Make your model run faster and use less memory.

02
Workload → hardware

Compilers & runtimes

Optimize MLIR/IREE, TVM, and LLVM toolchains for GPUs and custom accelerators, with hardware or simulator evaluation.

03
Domain structure → machine code

Specialized languages & toolchains

Build a software foundation around your particular workload.

One workload. An agreed baseline. A measurable next step.

Discuss a pilot

Research in the open

Ideas with evidence.

All research
01 / Mathematical discovery

22 years

A mathematical record, broken.

A new rigorous lower bound for square-lattice self-avoiding walks, improving the 2004 record.

Explore the result
02 / AI performance

2–4.5×

Faster kernels for AI.

Selected batched matrix multiplications on NVIDIA L40S, measured against torch.bmm / cuBLAS.

See the benchmarks
11,000+Spherical-code record campaignverified records · Discrete geometry45Matrix-multiplication schemesrecords · Algebraic complexity

The research engine

Every experiment
informs the next.

Agents explore hypotheses, build tools, and test results. Shared memory carries the evidence and working methods into the next campaign.

How autoresearch works
ExploreBuildCheckRemember

Work with Numaro

Put research to work.

Start with your workload and a clear performance target.

Discuss a pilot