The AI industry got remarkably good at building models. It never got good at freeing them.
Every serious AI team has the same experience. You train a model. It works. Then you try to deploy it on different hardware — and the work fractures. We started yasp to close that gap.
Yesterday
One model.
Three chips.
Three toolchains.
Three teams.
Months per target.
CUDA → TensorRT → ONNX → Vitis
{ Each chip = rebuild from scratch }
Today
Any model.
Any chip.
One command.
Binary out.
Done.
PyTorch → yasp → binary for any target
{ One path. Every chip. }
January 2023. Munich. Two engineers. One bet.
The bet: AI agents can generate better GPU code than hand-tuned heuristics.
Stefan Krassin had built AI infrastructure for companies that couldn't choose their own hardware. Reza Rahimi had written the optimization code that kept them locked in. They saw the same gap from both sides: every chip on the market required its own hand-written, hardware-specific kernels.
Their bet was specific: that AI agents, trained on enough GPU code, could learn to generate optimized hardware-specific code better than hand-tuned heuristics. Not for one chip. For any chip.
They registered yasp.ai and spent the next two years feeding GPU architecture manuals to language models, compiling the output, running it on real chips, and measuring whether the generated code was actually faster.
No pitch deck. No website. No Twitter account. Just the loop: generate, test, profile, iterate.
2023
Founded
Munich, Germany
$5M
Seed funding
September 2025
25
Engineers and researchers
across three cities
3
Munich · Montréal
New York
Member of
Members of Linux Foundation · PyTorch Foundation · NVIDIA Inception · EDGE AI Foundation · HAI Foundation
Four things decided early, and haven't changed.
These principles have guided every decision since day one - from product architecture to how we work with customers.
You choose the hardware
Not the vendor. Not the ecosystem. Not the toolchain. You. The person with the model and the idea.
Portability and performance
Not one or the other. Your model runs on the chip you chose, optimized for that target specifically. Not lowest common denominator.
Proven, not projected.
Every result validated on real silicon. The performance you see is the performance you get.
No new dependencies
No new language. No new stack. No new vendor. yasp works with what you already have.
Some of the people building this
The kind of team that argues about register allocation over lunch.
The story so far
We're not done.
Join us in building the infrastructure that makes model deployment fast and hardware choice real.
Munich. Montréal. Remote.
