Daisytuner
Category: AI Infrastructure
Darmstadt-based deep-tech startup that automates AI model performance engineering across any chip using an open-source compiler (docc) and the Daisyflow inference stack. Daisytuner was founded in 2024. The company is led by Lukas Trümper. Based in Darmstadt, Germany. Team size: 11-50. Total funding raised: $8.0M. Latest round: Seed. Key investors include Elaia, LEA Ventures, Amadeus APEX Technology Fund, Seedcamp, Inari VC, Linden Capital, Angel Invest, Mario Götze.
- Founded
- 2024
- Headquarters
- Darmstadt, Germany
- Team size
- 11-50
- Total funding
- $8.0M
Value proposition
Run any AI model on any hardware at full speed without rewrites, custom kernels or a new language, by automating the performance engineering that is normally redone for each new chip.
Products and solutions
docc (open-source compiler turning PyTorch, NumPy and C/C++ into a dataflow IR and native code for CUDA, ROCm, OpenMP and others), Daisytuner Platform (continuous benchmarking, web IDE, Performance Database), Daisytuner Enterprise (self-hosted, with SLA), Daisyflow (inference stack for language, vision, speech and recommendation models, with an OpenAI-compatible API, run on the customer's infrastructure).
Unique value
Automatic, hardware-agnostic and adaptive: AI agents search for optimizations on real hardware, the compiler validates each one, and validated recipes go into a performance database that speeds up later workloads.
Target customer
Cloud providers with mixed-hardware clusters, enterprises that want control of their data and execution, and chip vendors seeking faster production integration.
Industries served
AI infrastructure and cloud, HPC, enterprise AI, semiconductors / chip vendors
Technology advantage
Self-learning compiler with a verified optimization database. Early results show Qwen 2.5 up to 18% faster than vLLM on AMD Instinct GPUs. Research roots at ETH Zurich (SPCL) and RWTH Aachen.
How they differentiate
Compiler-level automation with a validated, continuously growing optimization database; vendor-neutral support spanning NVIDIA, AMD, Intel, Tenstorrent and photonic chips; deployment inside the customer's own infrastructure with hands-on operation.
Main competitors
Modular (now being acquired by Qualcomm), vLLM (open-source inference engine), other US and European compiler/inference startups
Key partnerships
Tenstorrent and Q.ANT (chipmaker collaborations), ETH Zurich SPCL, HiPEAC (Technology Transfer Award 2024, 2027 partner), Strategic Commercial Board with Codesphere and IONOS executives, other chipmakers under NDA.
Major milestones
2024: founded, HiPEAC Technology Transfer Award with SPCL (Dec 2024). 2025: CGO 2025 presentation with ETH Zurich, pre-seed round and beta platform launch, FOSDEM 2026 talk. Apr 2026: 3rd place at T-Challenge finals in Bonn. Nov 2026: €6M seed round and Daisyflow launch.
Growth metrics
Team of about 12 people with a target of 20+ by early 2027; close to finalizing first enterprise deals; Daisyflow in production.
Market positioning
Vendor-neutral performance layer and inference stack for mixed-accelerator environments, with a European sovereignty angle.
Geographic focus
Europe (Germany) with a global focus, including the US
Patents and IP
Core patent reportedly approved in late 2026 (per Tech Funding News); docc compiler is open source under a BSD-style license.
About Lukas Trümper
PhD research on AI-driven compilers at ETH Zurich (SPCL); studied at RWTH Aachen University. Left academia to co-found Daisytuner in 2024.
Latest news about Daisytuner
More AI Infrastructure companies
Official website: https://daisytuner.com