Daisytuner Raises €6M Seed to Automate AI Model Tuning Across Nvidia and AMD Chips
Darmstadt startup Daisytuner raised €6 million in seed funding and launched Daisyflow to compile and tune AI models across Nvidia, AMD, and Intel chips.

By Jin, Founder & Startups
, updated 4 min read
German compiler startup Daisytuner has raised €6 million in seed funding to automate how AI models are tuned for different silicon architectures, launching its Daisyflow inference engine into production alongside the round.
Elaia and LEA Ventures co-led the investment, joined by the Amadeus APEX Technology Fund and Seedcamp. Based in Darmstadt, the company builds software that removes the manual kernel engineering normally required to run large models efficiently on new accelerators.
Why compiler tuning bottlenecks AI deployments
Open-weight models have reached the point where enterprises can run capable systems inside their own data centres. As agent workflows multiply the number of inference calls per task, compute cost and latency depend heavily on how well model operators map onto the underlying processor.
Hardware diversity is compounding that problem. Alongside Nvidia GPUs, enterprises and European cloud operators are testing AMD Instinct accelerators, Intel chips, and newer custom silicon from companies such as Tenstorrent. Each architecture demands separate low-level optimisation, and minor model updates can break hand-tuned kernels.
"Demand for AI compute is growing faster than any one chip generation can serve," said Lukas Trümper, chief executive and co-founder of Daisytuner, in the Amadeus Capital Partners announcement. "The hardware to meet it is already out there, from established GPUs to a wave of new accelerators, but software keeps much of it on the sidelines."
How Daisyflow and DOCC work
Daisytuner pairs an open-source compiler called DOCC with an automated optimisation loop. The compiler translates code from standard frameworks such as PyTorch into native machine code for target accelerators.
In the background, software agents test candidate optimisations directly on physical hardware, verify numerical correctness, and save verified improvements to a shared performance database. Subsequent workloads draw on those stored patterns rather than starting from scratch.
Daisyflow packages that compiler into an enterprise inference server supporting language, vision, speech, and recommendation models behind standard APIs. In initial production tests reported by the company, Daisyflow ran Qwen 2.5 up to 18% faster than vLLM on AMD Instinct GPUs.
What the seed round funds
Daisytuner plans to use the €6 million to expand its engineering team, widen chip coverage in DOCC, and onboard enterprise customers and European hyperscalers onto Daisyflow. For teams locked into a single GPU vendor by CUDA tooling, automated cross-chip compilation offers a practical path to pricing leverage and multi-vendor resilience.
Sources
- 1Amadeus Capital Partners announcementamadeuscapital.com