THE ESSENTIALS IN 30 SECONDS
- Daisytuner raised a €6 million seed round led by Elaia and LEA Ventures, with Amadeus APEX Technology Fund and Seedcamp taking part.
- The start-up builds DOCC, an open-source compiler, and Daisyflow, an inference engine meant to run on customers' own hardware.
- Its claim that Daisyflow runs Qwen 2.5 up to 18% faster than vLLM on AMD Instinct GPUs is a company benchmark, not independently verified.
Daisytuner, a compiler start-up based in Darmstadt, Germany, has raised €6 million in a seed round, according to a report first published by EU-Startups and republished on 16 November by TechFlow. The round was led by the venture firm Elaia and LEA Ventures, with Amadeus APEX Technology Fund and Seedcamp also participating. The company had previously raised a €1 million pre-seed round led by LEA Partners, according to HiPEAC, the European computing research network.
What the company builds
Founded in 2024 by Lukas Trümper, its chief executive, and Adrian Schmitz, Daisytuner works on the software layer that sits between AI frameworks and the processors that run them. Its open-source compiler, DOCC, is described as self-learning: it takes code written for frameworks such as PyTorch and turns it into code tuned for a specific chip. The second product, Daisyflow, is an inference engine for language, vision, speech and recommendation models that customers can deploy on infrastructure they own.
Daisytuner's website lists support for Nvidia hardware through CUDA, AMD through ROCm, Tenstorrent through its Metalium software and the second generation of Q.ANT's photonic processor. On the same day as the funding news, the company posted an item on its site presenting Daisyflow as able to run models "on any chip".
Claims and plans
The headline performance figure comes from the company itself: Daisytuner says Daisyflow runs Alibaba's Qwen 2.5 model up to 18% faster than vLLM, a widely used open-source inference engine, on AMD Instinct GPUs. The test conditions were not detailed in the coverage, and no independent benchmark has been published. Daisytuner has not disclosed customer names or revenue.
The start-up says the money will go towards selling Daisyflow to enterprises, hiring engineers and extending its compiler to new chip architectures.
Why it matters
Most AI software is written and optimised first for Nvidia GPUs, which makes switching to AMD or to newer accelerators costly. A growing number of companies, from chipmakers' own software teams to start-ups, are trying to reduce that friction with portable compilers and inference stacks. Daisytuner's round is modest by AI standards, and its ability to keep pace with fast-moving hardware and models across several vendors remains to be shown in production deployments.
