Back to News
ProductAI Understanding briefing

Daisytuner raises €6 million seed to launch Daisyflow inference engine and open‑source compiler

Daisytuner announced a €6 million seed round and the release of Daisyflow, an AI inference engine that runs models on customers’ own hardware, backed by performance gains of up to 18 % over vLLM on AMD Instinct GPUs.

5 min readRead the linked source
Source-provided image accompanying Daisytuner raises €6 million seed to launch Daisyflow inference engine and open‑source compiler
Source referenceSource recorded
Publisher
unite.ai
Source type
Linked source — primary-source status has not been established.
ContextUnderstand this in 60 seconds

Key terms

Inference
The runtime phase where a trained model generates predictions or outputs.
Benchmark
A standardized test or dataset used to measure and compare model performance.
Latency
The time between sending a request and receiving the model's output.

What happened

Daisytuner secured €6 million in seed funding, co‑led by Elaia and LEA Ventures with participation from Amadeus APEX Technology Fund and Seedcamp. The capital will fund broader enterprise rollout of Daisyflow, the company’s new AI engine, and further development of its open‑source compiler, docc. Daisyflow enables on‑premise or private‑cloud serving of language, vision, speech and recommendation models across NVIDIA, AMD, Intel and Tenstorrent accelerators. The company reports that running Qwen 2.5 on AMD Instinct GPUs via Daisyflow is up to 18 % faster than the popular vLLM runtime. docc translates PyTorch, NumPy and C/C++ code into a data‑flow intermediate representation and emits native CUDA, ROCm or other accelerator code, automating many manual optimizations required when new models or chips appear. The hosted Daisytuner Platform provides continuous benchmarking, a browser‑based optimization editor and a performance database that feeds into Daisyflow’s production serving layer. Daisytuner will assess workloads, install the system, port models and stay involved in operations for customers that choose to run the stack on‑premise, in a data centre or private cloud.

On November 16, 2026 Daisytuner announced a €6 million seed round, co‑led by Elaia and LEA Ventures, with Amadeus APEX Technology Fund and Seedcamp also participating. The funding is earmarked for expanding the company’s enterprise footprint and further developing its compiler technology.

The same announcement introduced Daisyflow, an AI engine designed to run models inside a customer’s own infrastructure. Daisyflow supports language, vision, speech and recommendation workloads and can be deployed on‑premise, in a private cloud or in a customer‑owned data centre. It integrates with NVIDIA, AMD, Intel and Tenstorrent accelerators and offers OpenAI‑compatible APIs for language models, plus a ready‑to‑use chat interface.

Central to Daisyflow is docc, Daisytuner’s open‑source compiler released under a BSD license. docc converts PyTorch, NumPy and C/C++ code into a data‑flow intermediate representation, then generates native code for CUDA, ROCm and other back‑ends. The compiler automates many manual performance‑tuning steps that typically arise when new models or chips are introduced.

Daisytuner’s hosted platform provides continuous benchmarking on real hardware, a browser‑based optimization editor and a performance database that stores verified optimization recipes. These recipes can be reused automatically during later compilations, reducing the effort required to achieve high performance on new workloads.

The company reported that running the Qwen 2.5 model on AMD Instinct GPUs via Daisyflow achieved up to an 18 % speed improvement compared with the vLLM runtime, marking the first concrete performance data point for the new stack.

Source details: unite.ai

Why it matters

The launch of Daisyflow and its accompanying compiler addresses a growing demand for AI that can be kept within an organization’s own infrastructure, a need driven by data‑privacy regulations, cost‑control and the desire to avoid vendor lock‑in. By offering up to an 18 % speed advantage on AMD Instinct GPUs, Daisytuner demonstrates that its compiler‑driven approach can extract measurable performance gains without requiring bespoke engineering for each model‑hardware pairing. This could lower the total cost of ownership for enterprises that already own heterogeneous accelerator fleets, and it may accelerate adoption of newer, less‑optimised models as they become production‑ready. The open‑source BSD‑licensed docc compiler also provides a reusable building block for the broader community, potentially fostering a more competitive ecosystem around AI model serving and hardware integration. However, the performance claim is limited to a single (Qwen 2.5 on AMD Instinct) and has not been independently verified across other models, workloads or accelerator vendors. The breadth of the claimed speedup across the full range of supported model categories and hardware remains an open question, as does the pricing and licensing model for Daisyflow in enterprise deployments.

Enterprise AI adoption is increasingly constrained by data‑privacy, and cost considerations that favour on‑premise or private‑cloud . Daisyflow’s ability to keep requests, data and outputs within a customer’s own environment directly addresses these concerns.

Performance gains of up to 18 % on a major accelerator (AMD Instinct) suggest that the compiler‑driven approach can extract value from existing hardware investments, potentially lowering the total cost of ownership for organizations with heterogeneous fleets.

By open‑sourcing the docc compiler, Daisytuner contributes a reusable tool to the community, which could accelerate broader innovation in model‑to‑hardware compilation and reduce reliance on proprietary runtimes.

The seed funding indicates investor confidence in the market for on‑premise AI serving solutions, a segment that has been less highlighted compared with cloud‑native AI services. This may spur further capital into similar startups targeting the edge and private‑cloud space.

Nevertheless, the performance claim is limited to a single model and hardware configuration, and the company has not disclosed pricing, licensing terms or detailed roadmap for supporting additional accelerator vendors. These unknowns will affect the practical adoption and competitive positioning of Daisyflow.

Interactive Mechanism

Interactive Mechanism: How It Actually Works

Explore the underlying technology behind this development interactively.

Model Parameter Size:8B Parameters
VRAM Required5.5 GBGPU memory footprint
Target HardwareMacBook / Single GPUDeployment tier
Privacy100% Air-GappedLocal device capability
Core takeaway: Small, quantized models (3B–8B) now run directly inside smartphones and laptops with complete data privacy, while mammoth 400B+ models remain the domain of datacenter clusters.
Interactive Concept Check+10 Points
AI Models Explained Quiz

Which component of an AI application is the machine-learning model itself?

What to watch next

Key indicators to monitor include: (1) independent results that confirm or challenge the reported 18 % speedup across a wider set of models and hardware; (2) adoption signals from cloud providers or large enterprises that deploy Daisyflow on mixed‑hardware fleets; (3) the evolution of the docc compiler’s performance database and whether third‑party contributors add optimization recipes; and (4) pricing and licensing details that could affect the economic case for on‑premise versus public‑cloud alternatives.

Independent studies that test Daisyflow across a broader set of models (e.g., LLaMA, GPT‑4‑class) and hardware (NVIDIA, Intel, Tenstorrent) to validate the reported speedup.

Announcements of enterprise or cloud‑provider pilots that adopt Daisyflow, especially those that highlight cost savings or improvements.

Growth of the docc performance database through community contributions, which would demonstrate the scalability of the recipe‑based optimization approach.

Release of detailed pricing, licensing or SaaS models for Daisyflow, which will clarify the economic proposition for potential customers.

Feedback from early adopters regarding integration effort, operational overhead and the need for Daisytuner’s ongoing support.

Related guides & quizzes

Found this useful?