top of page

Daisytuner Announces Daisyflow: a High Performance AI Inference Engine for Any Hardware

Writer: Karan Bhatia
Karan Bhatia
3 minutes ago
2 min read

Daisytuner, automating the performance engineering that used to be rewritten for every new chip, led by Lukas Truemper and Adrian Schmitz, has released Daisyflow, a new AI inference engine powered by its self-learning compiler and patented optimization platform. The company has also raised €6 million in seed funding, co-led by Elaia and LEA Ventures, with participation from Amadeus APEX Technology Fund and Seedcamp.


Open-weight models have caught up with frontier capabilities. Today, any company can run state-of-the-art AI on its own hardware, enabled by an inference engine that adapts model execution to the underlying hardware for optimal performance.


Performance Engineering, Reinvented.


Every new model or chip makes previous optimizations obsolete, leaving performance engineering as a largely manual process. Built on more than a decade of compiler research, Daisytuner automates performance optimization for any hardware.


Its open-source compiler, docc, translates PyTorch, NumPy and C/C++ into native code for CUDA, ROCm and emerging architectures. AI agents continuously discover, benchmark, and verify optimizations, building a performance database that makes subsequent workloads faster.


Daisyflow brings these capabilities into production, serving language, vision, speech and recommendation models through an optimized inference stack with OpenAI-compatible APIs.


Initial results show Qwen 2.5 running up to 18% faster than industry-standard vLLM on AMD Instinct GPUs.


The platform is available as a hosted solution or through Daisytuner Enterprise, a self-hosted offering with an SLA.


A World Where Progress Runs at Full Speed.


The vision is to accelerate AI adoption across the ecosystem:


  • Cloud providers: Automating thousands of optimizations across diverse hardware fleets.

  • Enterprises: Enabling complete control over data and execution.

  • Hardware vendors: Accelerating chip deployment and time to production.


New models and hardware should advance without being held back by manual integration.


The team includes researchers and engineers from leading institutions such as ETH Zurich, TU Munich, RWTH Aachen, TU Darmstadt, and Lawrence Livermore National Laboratory.

bottom of page