Daisytuner Raises €6 Million And Launches Daisyflow AI Inference Engine

Daisytuner has raised €6 million in seed financing and launched Daisyflow, an AI inference engine designed to improve the performance of artificial intelligence models across different computing chips. The financing round was co-led by Elaia and LEA Ventures, with participation from APEX Ventures and Seedcamp.

The company is developing a self-learning compiler platform that uses AI agents to identify and test performance optimizations directly on hardware. The technology is intended to help organizations run AI models more efficiently while reducing the engineering work traditionally required to optimize software for different processors.

The launch comes as demand for AI inference continues to increase, with businesses deploying increasingly sophisticated models across a growing range of applications.

Inference refers to the process of running a trained AI model to generate predictions, responses, or other outputs. As organizations scale their use of AI, inference performance becomes an important factor in determining computing costs, application responsiveness, and infrastructure efficiency.

Unlike model training, which involves developing or refining an AI system, inference occurs whenever the deployed model processes new requests. Improvements in inference efficiency can therefore have a significant cumulative impact for businesses operating AI applications at scale.

One of the challenges facing AI infrastructure providers is the diversity of computing hardware available for running models.

Different processors have different architectures, memory systems, and computational capabilities. Software that performs efficiently on one type of chip may require additional optimization to achieve comparable results on another.

This creates an ongoing challenge for developers, who must adapt applications as new AI models and hardware platforms enter the market.

Daisytuner is seeking to address this complexity through Daisyflow, which automates portions of the optimization process.

Self-Learning Compiler Platform

Daisyflow uses AI agents to explore potential optimizations and evaluate their performance directly on the computing hardware where an AI model will run.

Rather than relying exclusively on manually developed optimization techniques, the platform tests different approaches to identify configurations that can improve execution speed.

Its compiler checks whether each proposed optimization produces correct results before accepting it.

This verification process is an important component of the technology because performance improvements must not come at the expense of computational correctness.

Once a successful optimization has been identified and verified, Daisyflow stores the result for future use.

Over time, this process is intended to build a collection of validated optimizations that can support additional workloads and reduce repeated engineering efforts.

The approach combines automated experimentation, performance measurement, and compiler-based verification.

Daisytuner’s objective is to make AI inference optimization more adaptive as computing hardware and model architectures evolve.

Traditional optimization processes can require specialized engineering expertise and substantial testing whenever organizations introduce new models, processors, or infrastructure configurations.

By automating more of that work, Daisyflow is designed to reduce the amount of manual intervention needed to improve AI performance.

The platform’s self-learning capabilities could also help organizations adapt more quickly to changes in the AI hardware ecosystem.

Performance Improvements On AMD Graphics Processors

Daisytuner reported that the Qwen 2.5 AI model can run up to 18% faster using Daisyflow than with vLLM, an established open-source inference serving engine, on AMD Instinct graphics processors.

The benchmark provides an early example of the performance improvements the company is targeting.

However, the reported result applies to the specific model, workload, and hardware configuration evaluated by Daisytuner.

Performance gains may differ depending on factors such as model architecture, processor characteristics, memory requirements, batch sizes, and inference settings.

The result should therefore not be interpreted as a universal 18% improvement across all AI applications or computing systems.

Even so, the comparison illustrates the potential importance of automated inference optimization as organizations look for ways to improve the utilization of expensive AI infrastructure.

Faster inference execution can help reduce the computing resources required to process certain workloads or increase the volume of requests that existing infrastructure can support.

For companies operating large-scale AI services, such improvements may contribute to lower infrastructure costs and better application performance.

Expanding Commercial Deployment

Daisytuner plans to use the €6 million seed financing to continue developing Daisyflow and expand its commercial deployment.

The investment will support the company’s efforts to strengthen its compiler technology, improve its automated optimization capabilities, and make the platform available to a broader range of AI infrastructure users.