Discoveries that come with their own proof.

The vision

For all of history, discovery has moved at the speed of a few people thinking.

Most ideas are wrong, and checking them is slow. So most ideas are never tried at all.

We're building a lab that tries them all, checks its own work, and shows you the proof.

How the lab works

Propose. Check. Keep.

Propose

Models suggest thousands of candidate answers: a construction, a faster algorithm, a better design.

Check

Every candidate passes through an automatic checker: a Lean proof, a test suite and benchmark, or a physics simulation.

Keep

What passes locks into place and becomes the starting point for the next round. What fails falls away, and is still recorded.

Where we're looking · Mathematics

Records that stood for decades.

Packings, constructions and Erdős problems with a published best-known answer. Then open questions nobody has settled, with proofs formalised in Lean.

Checker: code and Lean

Where we're looking · Algorithms

Faster ways to compute everything.

Kernels, scheduling and compression, measured against the best public implementations. A faster algorithm saves time on every machine that runs it.

Checker: tests and benchmarks

Where we're looking · Physics and engineering

Designs no one has drawn yet.

Antennas, optical structures and more, judged by open simulators first, then built and tested.

Checker: simulation, then prototype

Where we're looking · Inventions

A lab that never stops inventing.

Our own models, trained on every verified attempt, run the loop full time and turn results into inventions people can use.

Checker: whatever the claim needs

Every field

Then we take it everywhere.

Once the lab is big, the same loop goes to work on the hardest problems in every field. Each project gets its own way to check results: lab tests, field trials or real experiments with partners.

• Talking to animalsDecoding whales, birds and elephants, checked by playing sounds back with field researchers.
• Reading lost languagesScripts nobody can read yet, like Linear A and the Indus script.
• Designing proteinsEnzymes that break down plastic, tested in partner labs.
• New medicinesCandidates for diseases that get too little attention.
• Batteries and solarNew materials, simulated first, then made and measured.
• Fusion and climatePlasma control for fusion and materials that capture carbon.
• BrainsHow memory and language are stored.
• Chips and robotsAI-designed circuits and robots that learn new skills fast.

Later · after the core lab proves itself

Our own models

From small specialists to flagship models.

Every attempt the engine makes, passed or failed, becomes training data. That's how Quantavane builds its own models, step by step.

1. SpecialistsHundreds of small models, each the best at one job.
2. Our own modelsTrained from scratch on what the lab has learned.
3. Flagship modelsGeneral models at the frontier, built by Quantavane.

Our sister companies

Built to power the companies around it.

Quantavane's models run inside the companies it grows up with. A model only replaces what they use today when it measurably wins.

GIMBIX

Cuts what companies pay for AI.

Cirrun

Compute that stops when your code does. Its users get Quantavane models for free, and its machines help train them.

TraceAxiom

Checks what AI agents and code actually do.

CanliCapital

Financial research that separates real findings from luck.

How we work

Proof over claims.

If it can't be checked, we don't announce it.

Every result ships with the code that verifies it.

Previous record holders, collaborators and anyone who beats us are named.

Tools, benchmarks and attempts are open by default.

The next ten years

From the first record to the world's science.

1. 2027

   Prove the engine

   First published records beaten, the engine released as open source, first competitions.

2. 2028

   Get known

   Algorithm and physics results, first papers and the Open Problems Benchmark.

3. 2029

   Build the team

   Full-time researchers, hundreds of specialised models, first industry models.

4. 2030

   Our own models

   The first Quantavane model trained from scratch. Biotech begins.

5. 2031

   Many branches

   Materials, energy and talking to animals.

6. 2032

   An institute

   Research hubs and moonshots across five or more fields.

7. 2034

   Frontier

   Flagship models at the frontier.

8. 2036

   The engine for the world's science

   The leading lab for verified discovery.

Each stage starts when the last one has real results. Dates are targets.

Where this goes

The engine for the world's science.

Companies, universities and governments send it their hardest problems. It sends back answers with proof, in every field.

People

Started by one person. Built with many.

Quantavane was founded by Arhan Canli, who also builds GIMBIX, Cirrun, TraceAxiom and CanliCapital.

Competitions

We compete as Quantavane on Kaggle and in research challenges, and publish how we did, losses included.

Young researchers

Students anywhere can work on real problems with the lab and be credited on what they find.

Researchers

As the lab grows, we'll hire top researchers. Until then, send us a problem.