Darmstadt startup Daisytuner raises €6 million for AI inference on any chip

Darmstadt-based DeepTech startup Daisytuner has closed a €6 million seed round and launched Daisyflow, an inference engine meant to run AI models efficiently on whatever chips a customer uses. The company announced both on November 16, 2026.
Funding round and backers
Elaia and LEA Ventures, two European venture capital firms, co-led the seed round. The Amadeus APEX Technology Fund and Seedcamp also took part.
According to the company, the new capital will go towards bringing Daisyflow to a wider group of enterprises, expanding the team and developing its compiler technology further.
What Daisyflow is designed to do
Daisyflow serves AI models inside a customer's own infrastructure and tailors their execution to the hardware at hand. Built on open frameworks, it serves language, vision, speech and recommendation models through interfaces engineering teams already use, and includes a ready-made chat interface.
Requests are directed to the right models, batched and scheduled so the hardware stays busy, and multi-step and agent workflows are coordinated. Because it can run on a customer's own systems, companies keep control of their data and costs without first assembling large in-house engineering teams.
Compiler automation behind the product
The product rests on docc, Daisytuner's open-source compiler. It translates PyTorch, NumPy and C/C++ code into a dataflow representation and generates native code for NVIDIA's CUDA and AMD's ROCm platforms as well as newer architectures. The compiler is published under the new BSD licence.
AI agents look for optimisations around the clock and benchmark them on real hardware. Each result is verified for correctness before proven recipes are stored in a performance database, from which they can be matched to similar code in later compilations. The company describes this as a self-learning compiler combined with a patented optimisation platform.
Early performance figures and remaining challenges
Daisytuner reports that early results show Qwen 2.5 running up to 18% faster than vLLM on AMD Instinct GPUs. The company still needs to extend this early AMD result across more model types and accelerator vendors.
The startup says manual performance engineering has long been slow work carried out by a handful of frontier labs, and that new chips or model generations can invalidate earlier optimisations. Daisytuner presents automation as its answer to that problem.
Open questions
- Will the 18% speed-up for Qwen 2.5 on AMD hardware hold across other models and accelerator vendors?
- How quickly can Daisyflow reach more enterprises with the new funding?