· Funding· 日本語で読む
Germany's Daisytuner raises €6 million to optimize AI models across chips, launches inference engine
Daisytuner raised €6 million in seed funding, eu-startups.com reported. The round was co-led by Elaia and LEA Ventures, with participation from Amadeus APEX Technology Fund and Seedcamp. The company launched Daisyflow, an AI inference engine designed to improve execution efficiency across multiple chips.
What happened
eu-startups.com reported on November 16, 2026, that Daisytuner, based in Darmstadt, Germany, raised €6 million in seed funding. The round was co-led by Elaia and LEA Ventures, with Amadeus APEX Technology Fund and Seedcamp participating. Daisytuner also launched Daisyflow, an inference engine designed to run AI models efficiently on different hardware.
Why it matters
Daisytuner aims to use a compiler and inference engine to handle the process of tuning AI models for specific chips. For companies comparing and operating multiple accelerators, a practical question is whether the technology can reduce the amount of chip-specific code adjustment required. The reported result of up to 18% for Qwen 2.5 is an early result on AMD Instinct GPUs; disclosure of the testing conditions and validation on other models and chips will help determine how broadly it applies.
Analysis
The focus of this development is not only the €6 million in funding, but also the launch of an inference engine at the same time. The eu-startups.com article describes Daisytuner's work as a compiler and optimization platform that automates AI model performance tuning. The company aims to handle in software the process of adapting AI models to chips, and the business will depend on how much it can reduce the development burden of supporting individual accelerators. However, the materials provided do not specify how the funding will be allocated or the full range of supported chips.
The company cites NVIDIA and AMD GPUs as examples of target hardware. This can be understood not merely as running inference on multiple chips, but as combining a compiler that adapts model code to each chip with Daisyflow, which runs models during inference. However, being able to run the same model on different chips is a separate question from whether those chips are viable options in production in terms of performance, cost and power consumption. An article published by Amadeus Capital reports an early result showing Qwen 2.5 running up to 18% faster on an AMD Instinct GPU, but the materials do not provide the comparison conditions or test environment. That result alone is not enough to assess the effect across a wider range of models or applications.
We reported on leaderobot.com's limited sale of the five-fingered “XHAND 1 Pro-U” for NVIDIA GEAR DexUMI, which covered an NVIDIA robotics-related configuration. Daisytuner's announcement concerns the software layer supporting AI inference, not robotic hardware. No direct connection between the two has been confirmed, but when considering the computing infrastructure used in robotics development and operations, the relationship between supported hardware and software also needs to be assessed individually. Going forward, the areas to watch include Daisyflow's range of supported chips and models, benchmark conditions, latency and costs in production, and the status of its enterprise offering. The funding and product information here is based on secondary reporting; readers should consult the original sources for details.
Sources
Translated from the Japanese original: /news/doku-daisytuner-ai-moderu-no-chippu-kan-saitekika-wo-mezashi-39a82c. The Japanese version is authoritative if they differ.