Daisytuner Raises €6 Million Seed Round to Expand Daisyflow and Compiler Technology
On November 16, 2026, Daisytuner announced the completion of a €6 million seed funding round, co-led by Elaia and LEA Ventures, with participation from Amadeus APEX Technology Fund and Seedcamp, and simultaneously released the AI inference engine Daisyflow. Daisyflow is deployed in customers' own infrastructure, supports language, vision, speech, and recommendation models, and offers optional NVIDIA, AMD, Intel, and Tenstorrent accelerators. Early results show that Qwen 2.5 runs up to 18% faster on AMD Instinct GPUs than vLLM.
Body
Daisytuner announced on November 16, 2026 that it had raised €6 million in seed funding and released Daisyflow, an AI inference engine that serves models inside customers’ own infrastructure and adapts their execution to different hardware.
Elaia and LEA Ventures co-led the round, with participation from Amadeus APEX Technology Fund and Seedcamp. Daisytuner said the capital will be used to bring Daisyflow to more enterprises, grow its team and expand its compiler technology.
Daisytuner reports that its first results show Qwen 2.5 running up to 18% faster than vLLM on AMD Instinct GPUs. The result gives the launch a specific performance data point as the company presents Daisyflow for a broader range of model categories and accelerator suppliers.
Compiler Automation Behind Daisyflow
At the center of the technology is docc, Daisytuner’s open-source compiler. It translates PyTorch, NumPy and C/C++ into a dataflow intermediate representation, a format that makes data movement, parallelism and computation explicit, then generates native code for NVIDIA’s CUDA and AMD’s ROCm platforms and for novel architectures. This addresses the manual performance work that can arise when a new model or chip disrupts existing optimizations.
AI agents search for optimizations and benchmark them on real hardware. docc verifies each optimization for correctness before use, and proven recipes enter a performance database. During later compilation, the compiler can match those recipes to similar code, allowing knowledge from earlier workloads to inform new optimization work.
The hosted Daisytuner Platform turns this into a developer workflow with continuous benchmarking on GitHub pushes, a browser-based optimization editor and the performance database. Users can inspect hotspots measured on real hardware, create optimization recipes and submit the results as pull requests. Daisyflow builds on that optimization layer by adding production model serving.
The docc repository publishes the compiler under the new BSD license. Its compatibility matrix lists Linux support for C/C++, Python and PyTorch frontends with OpenMP, CUDA and ROCm backends. The PyTorch frontend supports inference only rather than training.
Customer-Controlled Deployment
Daisyflow serves language, vision, speech and recommendation models. Its inference layer routes requests to models, batches and schedules work for hardware utilization, and coordinates multi-step and agent workflows. Applications connect through standard APIs, with OpenAI compatibility for language models, and a ready-to-use chat interface is included.
The Daisyflow product page lists deployment on premises, in a customer’s data center or in a private cloud, with NVIDIA, AMD, Intel and Tenstorrent as accelerator options. Requests, data and outputs remain within the customer’s infrastructure. Daisytuner assesses workloads, hardware and costs, installs the system, ports models to the selected accelerators and remains involved in operation and tuning.
Daisytuner is targeting cloud providers with mixed hardware fleets, enterprises that want control over data and execution, and hardware vendors seeking faster production integration. The deployment model connects those audiences: customers choose or own the infrastructure, while Daisytuner supplies the optimization and serving layers and works with them on operations.
The funding supports both sides of that offer: wider enterprise deployment and a compiler system intended to reduce manual integration as new models and chips arrive. The central challenge is breadth, extending the early AMD performance result across the model types, accelerator vendors and customer environments Daisyflow is designed to support.