Skip to main content
v2026.11,986 entries · CC-BY 4.0
CASRAIAsk CASRAI · AI assistant for researchAsk a research question. Answers come with their sources.Funder rules, compliance, methods and publishing. Answers link the sources they rest on.

Editorial · CASRAI · AI and ML research outputs

Daisytuner raises €6m and releases Daisyflow AI inference engine

Daisytuner announces €6 million in seed funding and Daisyflow, an AI inference engine designed to run models inside customers’ infrastructure.

Published 16 Nov 2026· 4 minute read

Ask CASRAI · free to try

Ask about this story

2 free questions a day, no account.

An AI assistant for research: funder rules, compliance, methods and publishing. It links the sources behind its answers and labels answers that come from a web search.

Answers draw on CASRAI's guides and dictionary plus the federal and funder documents we index: Federal Register, Grants.gov, Regulations.gov and UKRI.

Works on this site and inside Claude, Cursor and the AI tools you already use.

Everything CASRAI publishes — this page, the dictionary, the guides and the news — stays open to read, with no account and no card.

CASRAI is the reference for research administration — bookmark it for the next question.

Daisytuner announced €6 million in seed funding and the release of its Daisyflow AI inference engine on November 16, 2026. The Darmstadt company says the system runs models inside customers’ infrastructure and optimizes execution for different chips. For university computing teams, the development puts software portability and locally operated AI services alongside the question of which accelerators to buy.

A seed round for compiler technology and model serving

In its November 16 announcement, Daisytuner named Elaia and LEA Ventures as the round’s co-leads, with Amadeus APEX Technology Fund and Seedcamp participating. It plans to use the capital to expand enterprise adoption of Daisyflow, grow its team and develop its compiler technology. These are the company’s deployment and development plans, rather than announced university awards or public research grants.

Tech.eu’s report separately confirms the funding and identifies the focus as inference: running an existing AI model rather than training one. That distinction matters for research offices evaluating an institutional AI service. A system for serving language or vision models addresses a different operational task from infrastructure for training a new model.

What Daisyflow adds to the compiler

The company’s Daisyflow product description separates the customer’s accelerators and applications from two supplied layers: Daisytuner Enterprise for optimization, and an inference stack for serving models. It lists on-premises, data-center and private-cloud deployments, with NVIDIA, AMD, Intel and Tenstorrent among its accelerator options. Those are the vendor’s stated options, rather than a claim that every model has been independently tested on every listed device.

Daisyflow provides request routing, batching and scheduling, and supports multi-step workflows. Daisytuner lists language, vision, speech and recommendation workloads, with OpenAI-compatible interfaces for language models. The company says requests, data and outputs remain inside the customer’s infrastructure. It also describes an operational service: assessing workloads, deploying the stack, porting models and continuing to tune the installation with the customer.

The underlying docc compiler has a different role. According to Daisytuner’s technical explanation, it represents computation and data movement in an intermediate form, checks proposed optimization recipes and stores accepted recipes in a performance database. Agents search for optimizations, while compilation can reuse recipes for similar code. The company also describes continuous benchmarking from GitHub pushes and a browser-based editor for reviewing transformations.

The benchmark is specific, and the compatibility limits matter

Daisytuner reports that Qwen 2.5 ran up to 18% faster than vLLM on AMD Instinct GPUs. Both the company announcement and Tech.eu attribute that result to the company. It should therefore be treated as a vendor benchmark for a named model and hardware family. It does not establish an equivalent gain for a university’s other models, batch sizes or accelerator configurations.

The docc repository supplies a more concrete compatibility reference. Its matrix lists Linux support for C/C++, Python and PyTorch with OpenMP, CUDA and ROCm backends, while several macOS combinations remain works in progress. Its PyTorch instructions explicitly limit support to inference. The repository also publishes the compiler under the BSD-3-Clause license. These details help distinguish the inspectable compiler from the broader commercial deployment offer.

Research roots and the institutional decisions ahead

Daisytuner’s company profile traces its origins to research at ETH Zurich and RWTH Aachen. It identifies Lukas Truemper as founder and CEO and Adrian Schmitz as co-founder. The new funding consequently provides a research-commercialisation example as well as an infrastructure announcement: compiler research is being developed into a service for operating AI workloads.

In an interview with Tech Funding News, Trümper described demand for moving code across accelerator suppliers. The outlet reports that most funding will support headcount, with additional spending on hardware for testing and benchmarking. Its interview also conveys investors’ emphasis on compute efficiency and technology sovereignty. Those views explain the commercial rationale; they are stakeholder assessments rather than independently measured product outcomes.

Editorial interpretation: for universities, the relevant evaluation is whether a local inference service works on their actual workloads and fits their operating arrangements. Keeping data within an institution’s infrastructure and having vendor support are separate questions from reproducing a speed claim. A pilot could combine research software code review with comparative performance measurement, recording model versions, hardware and workload conditions before wider deployment.

The company’s next stated steps are broader enterprise deployment, team growth and compiler development. Its announcement sets no dated rollout milestones for those steps. For research computing groups, the immediate evidence to examine is therefore the released compiler’s compatibility information and a workload-specific assessment of the serving stack.

Sources

Referenced across the research world

University of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logoUniversity of Cambridge logoColumbia University logoCrossref logoUniversity of Edinburgh logoHarvard University logoUniversity of Oxford logoPrinceton University logoStanford School of Medicine logoUniversity College London logoORCID logo
  • University of Cambridge logo
  • Columbia University logo
  • Crossref logo
  • University of Edinburgh logo
  • Harvard University logo
  • University of Oxford logo
  • Princeton University logo
  • Stanford School of Medicine logo
  • University College London logo
  • ORCID logo

View CASRAI adoption →