Daisytuner Raises €6M Seed to Expand Daisyflow AI Inference
Adrian Schmitz brought object detection onto a photonic processor at Deutsche Telekom's Bonn headquarters in April. In Daisytuner's account of the demonstration, the model reached Q.ANT hardware through the company's compiler stack without changes to its PyTorch code. The commercial possibility starts there: give an application another place to run without handing its developers another rewrite.
Daisytuner, based in Darmstadt, Germany, is taking that compiler work further into enterprise AI with a €6M Seed round announced on November 16, 2026. Elaia and LEA Ventures co-led the financing, with Amadeus APEX Technology Fund and Seedcamp participating. Alongside the funding, the company released Daisyflow, an inference engine that brings its optimization technology into the business of serving models on customer infrastructure.
The founders are selling access to a scarce kind of engineering knowledge: how to make a particular workload use a particular machine well. Turning that knowledge into something repeatable could change the calculation behind an enterprise's next accelerator purchase. The financing will support more enterprise deployments, team growth and further compiler development, according to the company.
From a hardware demonstration to an operating service
Daisyflow expands the assignment Daisytuner is willing to take. The company's earlier compiler work concerns how an application becomes executable code for a processor; its inference offering adds the surrounding service needed to run trained AI models. Serving requests, grouping work into batches and routing requests to models all become part of the product conversation.
That matters to the person buying the system. Choosing hardware can also mean committing engineers to integrations, tuning and maintenance that continue after the equipment arrives. A processor with an attractive price still has to fit the models an organization uses and the interfaces its applications expect. Software support becomes part of the purchasing decision before any performance comparison reaches a spreadsheet.
Daisytuner describes Daisyflow as a stack installed on customer-owned infrastructure, with the company helping operate and optimize it. Standard interfaces let applications connect to the service, while the offering covers language, vision, speech and recommendation workloads. The proposition combines control over the environment with outside specialist involvement in keeping it useful.
Those responsibilities give the funding an operating purpose. Expanding enterprise deployments means putting compiler knowledge into environments where someone expects the service to keep working as software changes. An optimization that looks convincing in a demonstration has to become part of a maintained relationship with the team running the application.
What the compiler keeps from the last assignment
Daisytuner's technical explanation describes a compiler called docc, a database of optimization recipes and measurements collected on real hardware. The compiler represents computation and data movement explicitly, allowing transformations to be checked before the resulting recipes are reused. Agents search for improvements, while stored results give later work a starting point.
The economic interest lies in that accumulation. A specialist can spend time finding a better way to execute a workload, yet a new model or processor can send another team back through similar work. A database that makes useful prior work discoverable offers a way to carry engineering effort across assignments. The quality and applicability of those recipes therefore matter alongside the search process that produces them.
This is the commercial interpretation of Daisytuner's design, rather than a measured claim about customer savings. The company has not supplied a verified customer cost model for this report. Buyers would still need to assess the workloads, service commitments and integration work relevant to their own deployment.
The product lineup gives different teams different entry points. Daisytuner Platform is the hosted tooling offering, Enterprise provides a self-hosted environment with support commitments, and Daisyflow adds the inference stack and operational involvement. Keeping those layers distinct helps explain what an enterprise is buying beyond access to an open-source compiler.
The founders and the financing behind the move
Founder and CEO Lukas Trümper and co-founder Adrian Schmitz connect the business to its research origins. Daisytuner identifies ETH Zurich and RWTH Aachen among those roots; Trümper's company biography links his founding work to AI-driven compiler research at ETH Zurich. Schmitz's stated focus is GPU optimization.
The new €6M round also follows an earlier financing. HiPEAC documented a €1M pre-seed led by LEA Partners when Daisytuner launched its beta benchmarking product. That history puts the current announcement in a progression from helping developers inspect software performance to taking on more of the execution environment around AI models.
The November announcement names Elaia and LEA Ventures as the current co-leads, using the investor names attached to this round. Seedcamp also lists Daisytuner in its company portfolio. The disclosed €6M is the seed financing amount; the available material does not establish a company-confirmed lifetime funding total or a valuation.
Daisytuner's open positions span GPU performance, compiler engineering, product, distributed systems, infrastructure and full-stack development. Those roles describe the breadth of work behind the expansion: compiler research has to meet product delivery and the daily demands of customer environments.
Hardware choice now includes the maintenance conversation
Daisytuner's technical record reaches beyond language models. In January, the company described running OpenFOAM on Tenstorrent hardware, extending its compiler approach into an established simulation application. The April Q.ANT demonstration supplied another example on a different computing architecture. These are documented technical projects, rather than evidence of named paying-customer contracts.
The Daisyflow launch adds a more specific performance claim: Daisytuner reports that Qwen 2.5 runs up to 18% faster than vLLM on AMD Instinct GPUs. The comparison belongs to that model and hardware setting, and independent reproduction was not established for this report. It provides an initial reference point while leaving buyers to assess the workloads they actually intend to serve.
The market consequence is a wider conversation about who carries the work needed to make another chip useful. Hardware vendors need applications that can reach their processors; enterprise teams need a manageable path from a model they want to use to a service they can run. Daisytuner is trying to earn a place between those parties by retaining optimization knowledge and taking responsibility for more of the deployment.
For Trümper and Schmitz, enterprise expansion will put that proposition into more customer environments. A model update, a different accelerator or a new workload gives the compiler another assignment, and gives the team operating Daisyflow another reason to return to the work already stored in its database.
AI Infrastructure funding, last 30 days
DevCuration's funding database tracked 32 AI Infrastructure rounds totaling $26.7B in disclosed capital over the past 30 days. Recent deals we covered:
- Zeromatter Discloses $160M Raised at $1.1B Valuation$160M · Nov 16
- Rowan Adds $1.545B in Credit for Hyperscale Data CentersDebt · $1.545B · Nov 14
- Mirendil Raises $1B Series A for Agent-Led AI ResearchSeries A · $1B · Nov 13
- Mount Thor Raises $45M for Apple-Based AI InfrastructureSeed and Series A · $45M · Nov 13
- Plurai Raises $13M Seed for AI Agent SimulationSeed · $13M · Nov 13
Frequently Asked Questions
How does Daisyflow differ from Daisytuner’s compiler tools?
Daisytuner’s compiler and optimization platform help prepare and tune code for hardware. Daisyflow adds inference serving, routing and batching, with Daisytuner helping install and operate the stack on customer infrastructure.
Why can compiler optimization affect an enterprise’s hardware choices?
Applications need software support and tuning to use a processor efficiently. Reusing validated optimization recipes may reduce repeated engineering work when teams adopt different hardware; actual benefits depend on the workload and deployment.
What does the reported 18% performance improvement cover?
Daisytuner reports Qwen 2.5 running up to 18% faster than vLLM on AMD Instinct GPUs. This is a company-reported comparison for a specific model and hardware setting, rather than an independently reproduced result across all supported workloads.
What will the seed financing support?
Daisytuner says the €6M Seed round will support more enterprise deployments of Daisyflow, team growth and further compiler technology development. Elaia and LEA Ventures co-led the round, with Amadeus APEX Technology Fund and Seedcamp participating.
Where the Money Moved
The intelligence briefing of the innovation economy. Funding, M&A, debt and fund closes, read as market signal rather than deal announcements.
Subscribe to Where the Money Moved




