Daisytuner has raised seed funding in a round co-led by European venture capital firms Elaia and LEA Ventures. Amadeus APEX Technology Fund and Seedcamp also participated.
Alongside the fundraise, the 2024-founded startup launched Daisyflow, a high-performance AI inference engine. The platform aims to run AI models efficiently across different hardware architectures.
The funding comes as AI companies increasingly focus on optimising existing computing infrastructure. Consequently, improving chip-level performance has become important for addressing growing AI compute requirements.
Daisytuner targets AI hardware optimisation bottleneck
Founded in 2024 by Lukas Trümper and Adrian Schmitz, Daisytuner automates AI model performance engineering. Its technology combines a self-learning compiler with a patented optimisation platform.
The team brings more than a decade of research experience. Its backgrounds include ETH Zurich, TU Munich, RWTH Aachen, TU Darmstadt and Lawrence Livermore National Laboratory.
The startup’s open-source compiler, DOCC, translates code from frameworks such as PyTorch into native code. It supports different target chips, including GPUs from NVIDIA and AMD, alongside emerging architectures.
AI agents continuously test optimisations on real hardware. Furthermore, the system stores verified results in a performance database for reuse across similar workloads.
This approach aims to reduce reliance on manual performance engineering. Traditionally, specialised teams have handled this time-consuming process.
However, growing hardware diversity creates additional challenges. New chip and GPU generations, as well as different AI models, can require fresh optimisation work.
“Compute is the scarcest resource in AI, and most of it currently sits behind a single vendor’s software stack,” said Sébastien Lefebvre, Partner at Elaia.
He said Daisytuner allows the same code to run across different silicon without requiring rewrites. Lefebvre added that the company’s compiler technology addresses an important industry bottleneck.
Daisyflow brings AI inference across hardware platforms
Daisyflow is designed as a production-ready inference engine for enterprise deployments. It supports language, vision, speech and recommendation models through familiar APIs.
The platform is built on open frameworks and can run within customers’ own infrastructure. Therefore, enterprises can retain control over their data and computing costs.
According to Daisytuner, its platform ran Qwen 2.5 up to 18 per cent faster than vLLM on AMD Instinct GPUs.
The company has made Daisyflow available through its Cloud Console. The platform combines the DOCC compiler, continuous hardware benchmarking and a web-based IDE.
It also includes a live performance database. The compiler automatically queries this database during compilation to identify relevant optimisation results.
Daisytuner plans enterprise expansion and chip support
The startup plans to use the fresh capital to expand Daisyflow’s enterprise reach. Additionally, it will grow its engineering team and support more emerging chip architectures.
Co-founder and CEO Trümper said demand for AI computing was growing faster than individual chip generations could meet. He said the company was working with chip vendors, European hyperscalers and enterprises.
The startup aims to help organisations use existing hardware more efficiently. It also wants to support new accelerators as they enter production.
With major AI companies exploring custom silicon, Daisytuner is targeting a broader role. It aims to provide software that improves performance across different chips and computing platforms.

