Funding & VC

Daisytuner launches Daisyflow, raising €6 million for AI inference

Daisytuner raised €6 million and launched Daisyflow, an AI inference compiler that automates hardware optimization. It reported Qwen 2.5 running up to 18% faster than vLLM on AMD Instinct.

By Nathan Brooks

4 min read

Daisytuner Raises €6 Million And Launches Daisyflow AI Inference Engine
Daisytuner Raises €6 Million And Launches Daisyflow AI Inference Engineblickpixel / pixabay

What's News

  • Daisytuner raised €6 million in seed financing co-led by Elaia and LEA Ventures.
  • APEX Ventures and Seedcamp participated in the seed round.
  • Daisyflow uses AI agents to test and verify inference optimizations on target hardware.
  • Daisytuner reported Qwen 2.5 running up to 18% faster than vLLM on AMD Instinct graphics processors.
  • Daisytuner said performance gains vary by model, processor, memory, batch size and inference settings.

By Nathan Brooks

Daisytuner raised €6 million in seed financing and launched Daisyflow, an AI inference engine designed to improve AI model performance across different computing chips. Elaia and LEA Ventures co-led the financing, with APEX Ventures and Seedcamp participating.

Daisytuner will use the seed capital to continue developing Daisyflow and expand commercial deployment. The company said the investment will strengthen its compiler technology, improve automated optimization and make the platform available to more AI infrastructure users.

Daisyflow learns from the hardware

Daisytuner describes Daisyflow as a self-learning compiler platform that uses AI agents to find and test performance optimizations on the hardware where a model runs. The agents explore possible approaches, measure execution speed and identify configurations that can improve performance. The company says this automation can reduce engineering work traditionally required to optimize software for different processors.

Rather than relying only on manual techniques, Daisyflow tests several optimization approaches on target hardware. Its compiler checks that each proposed optimization produces correct results before accepting it. Daisytuner says that safeguard prevents speed gains from coming at the expense of computational correctness.

Once the system verifies a successful optimization, it stores the result for later use. Over time, Daisytuner expects those experiments to build a collection of validated optimizations that can support additional workloads and reduce repeated engineering. The company describes the method as “automated experimentation, performance measurement, and compiler-based verification.”

That combination is intended to make inference optimization adapt as hardware and model architectures evolve. Daisytuner says conventional optimization can require specialized engineering expertise and substantial testing whenever teams introduce a new model, processor or infrastructure configuration. Automating more of that process should, in the company’s view, help organizations respond faster to changes in the AI hardware ecosystem.

The 18% result has narrow boundaries

Daisytuner reported that the Qwen 2.5 AI model can run “up to 18% faster” with Daisyflow than with vLLM, an established open-source inference serving engine, on AMD Instinct graphics processors. The company presented the benchmark as an early example of the performance improvements it aims to deliver, rather than a universal performance claim.

Daisytuner said performance gains may vary according to:

  • model architecture;
  • processor characteristics;
  • memory requirements;
  • batch sizes; and
  • inference settings.

The company explicitly said its result does not apply as an 18% improvement across all AI applications or computing systems. It applies only to the model, workload and hardware configuration Daisytuner evaluated.

What does this change for AI infrastructure teams?

Daisytuner says faster inference execution can reduce the computing resources needed for some workloads or let existing infrastructure support more requests. For companies operating large-scale AI services, the company says those improvements may lower infrastructure costs and improve application performance.

The practical change is less manual optimization, not a new model-training method. Daisytuner says traditional optimization demands specialized engineering and extensive testing whenever models, processors or configurations change. Daisyflow aims to automate more of that work, so teams can adapt applications as new models and hardware platforms arrive.

Faster execution also gives infrastructure providers more room to improve utilization of expensive AI hardware. Daisytuner frames the AMD Instinct benchmark as an early illustration of why automated optimization matters as organizations look for lower computing costs and better application responsiveness.

Inference is the expensive, repetitive phase

Daisytuner defines inference as running a trained AI model to generate predictions, responses or other outputs. Unlike model training, which develops or refines an AI system, inference occurs each time a deployed model processes a new request.

The company says inference efficiency determines computing costs, application responsiveness and infrastructure efficiency at scale. It also says improvements can have a significant cumulative impact for businesses operating AI applications. Demand for inference is rising as businesses deploy increasingly sophisticated models across a growing range of applications.

Hardware diversity makes that task harder. Different processors have different architectures, memory systems and computational capabilities. Software that performs efficiently on one chip may need additional optimization to deliver comparable results on another, and developers must adapt applications as new models and hardware platforms reach the market. Daisyflow targets that gap by automating parts of the optimization process.

Seed capital goes toward deployment

Daisytuner plans to deploy the €6 million across three priorities:

  • continued development of Daisyflow;
  • stronger compiler technology and more capable automated optimization; and
  • broader commercial availability for AI infrastructure users.

The round gives Daisytuner resources for both product development and a wider customer footprint. The company is betting that self-learning compilers can absorb repetitive, hardware-specific tuning while inference demand expands.

For infrastructure buyers, the near-term test is whether Daisyflow can reproduce gains under their own models, batch sizes and infrastructure settings. Daisytuner’s 18% benchmark offers an early data point, but the company has not claimed that result applies universally.

The financing gives Daisytuner a path from a promising compiler benchmark to wider commercial deployment. The decisive test will be whether that expansion delivers repeatable performance gains without sacrificing computational correctness.

Source: pulse2.com

Share this article:

More from Nathan Brooks

Nathan Brooks

Show full bio

News editor covering regulation, policy and their impact on entrepreneurship, startups and venture funding at Business Bearings.

2,460 articles

Related articles

« Previous articleNext article »