Press Releases

yasp was founded on a simple observation: many of the best ideas in AI never reach their potential because performance optimization is too painful and time-consuming. Over the past two years, the team focused on one question: how can AI automate the ugliest parts of making AI fast?

That work led to the Agentic AI Compiler — the intelligent core of the future AI stack, bridging the gap between AI model development and heterogeneous infrastructure without a single rewrite of code.

Now, we are ready for the next step.

We are opening the doors to Gaia and Kronos.

>> Claim your access key

Two packages. Same platform.

The compiler now ships as two packages, because finding performance and shipping it are different jobs.

Gaia: maximum throughput, multi-vendor. Agents profile your model on the target hardware, generate candidate kernels, evaluate them numerically against the PyTorch reference, and refine until both correctness and performance thresholds are met. Output: an optimized PyTorch model that drops straight into TorchServe, Triton, FastAPI, or any PyTorch-based serving framework. Targets Nvidia cloud silicon, AMD GPUs and AWS Trainium, emitting CUDA, HIP or Triton. Built for hyperscalers, hardware vendors and ML labs.

Kronos: the path to production. PyTorch model in, self-contained binary out. No PyTorch in production, compatible with any CUDA version, OS or runtime version, with custom kernels generated for any operation the vendor toolchain doesn’t support natively. Currently optimized for Nvidia edge and production silicon including Jetson Orin and Drive AGX. Built for robotics, autonomous vehicles, drones and edge AI.

Start with Gaia. Ship with Kronos. Same compiler underneath — and every number on both sides came off real hardware.

Why early access is exclusive

By working closely with a limited group of teams across geographies and industries, we put Gaia and Kronos through the ultimate pre-launch stress test — from experimental labs to large-scale production systems. We get to fine-tune our UI and documentation; you get the chance to try out the platform before everyone else.

Act now: slots are limited to ensure tight feedback loops and dedicated support.

>> Secure your spot

How it works

The good news is: joining the program is easy. Instead of pushing developers deeper into low-level tuning, yasp fits into existing workflows. You send your workload through the compiler, and it takes over the routine heavy lifting: profiling runs, auto-generating optimized kernels, and tuning them for the GPUs and cloud environments you already use.

The compiler observes how your models behave on real hardware and uses that feedback to iteratively generate and refine kernels tailored to your setup. There is no need to rewrite your stack or adopt a new framework. Change the target, recompile — and when the work is done, yasp hands back the artifact and gets out of the way. No persistent layer, no forced runtime, no new dependency.

Every verified kernel also feeds the next compilation. Future runs on the same hardware hit the cache and complete in seconds, so the system gets faster and cheaper the longer it runs.

Proven on real silicon

AI infrastructure is full of performance claims that are hard to compare and harder to trust, so every result below was measured on the actual target hardware and validated against the PyTorch reference:

• Up to 3× end-to-end on IBM Granite 4.0 versus torch.compile, with identical text outputs
• 6.25× on the Granite 4.0 Mamba layer — a novel architecture, optimized automatically
• 2.91× for MiniGPT on Azure’s AMD Radeon PRO V710, and 67% lower cost than an Nvidia A10
• 30% lower perception deployment cost for an autonomous trucking customer, with development cycles compressing from months to weeks

Crucially, these gains come without sacrificing accuracy or forcing compromises in model design. Where an operation can’t be optimized, it falls back to the reference implementation, so correctness is never traded for speed.

What early access users get

The practical effect is shorter iteration cycles and fewer infrastructure detours. Instead of pausing feature work to “go optimize,” teams keep shipping while the compiler refines performance in the background. What once took days of manual tuning starts to feel like a flag in the toolchain.

Early access participants unlock the full core experience:

• Gaia for maximum throughput across Nvidia, AMD and AWS Trainium
• Kronos for self-contained binaries on Nvidia edge and production silicon
• Automated profiling, kernel generation and numerical validation on real hardware
• KernelDB cache access, so repeat compilations return in seconds

Beyond performance, participants get first-hand support automating and optimizing their model tuning operations, and the ability to shape the roadmap together with the yasp team. This goes beyond controlled demos: sharing traces, benchmarks and edge cases helps define what “AI optimizing AI” should look like in real production settings.

How to get involved

If you have ever shipped an AI model thinking “this should be faster, but there’s no time to tune it properly,” this program is built for you and your team.

Apply for early access, get hands-on with Gaia and Kronos, and influence where the roadmap goes next.