Agentic Engineering & ToolingConference50min
Quixotic: Pure Java Inference at Native Speed
Quixotic brings AI to Java by running models in-process as pure Java, with no Python, ONNX, or native dependencies. It keeps sensitive data inside the JVM, compiles models into fast native binaries, and enables secure, portable AI for regulated industries.
talk.summaryAiDisclaimer
Alfonso² PeterssenQuixotic AI
talkDetail.whenAndWhere
Thursday, October 8, 11:50-12:40
TBA 6
talks.roomOccupancytalks.noOccupancyInfo
AI for Java, in Java.
AI sovereignty for the world's most trusted runtime.
What if AI on the JVM were just one Maven dependency away? AI in a jar.
What if AI on the JVM needed no Python, no Docker containers, no native wrappers, no external processes, no HTTP requests?
What if AI on the JVM ran at native speed on plain CPUs, and sometimes even outperformed native inference engines?
What if every component of AI on the JVM, from tokenization to model formats, quantization, inference algorithms and prompt caching, were a pure-Java building block?
What if AI on the JVM could not only read and write, but also hear and see?
What if AI on the JVM were completely free, open source and built on open standards?
This talk shows how inference is built using pure Java and how to make local AI run fast on ordinary CPUs. You'll leave knowing how modern LLMs actually work under the hood, how the Vector API and quantized matrix multiplication kernels close the gap with native engines, and how to add local inference to your own Java application with a single dependency.
Come for the live demos: Gemma 4, LFM 2.5+, gpt-oss, Qwen 3.5+, Ministral 3, Granite 4.1+ and Nemotron 3+ on a vanilla JVM, at native speed. Text, audio and vision demos included.
AI sovereignty for the world's most trusted runtime.
What if AI on the JVM were just one Maven dependency away? AI in a jar.
What if AI on the JVM needed no Python, no Docker containers, no native wrappers, no external processes, no HTTP requests?
What if AI on the JVM ran at native speed on plain CPUs, and sometimes even outperformed native inference engines?
What if every component of AI on the JVM, from tokenization to model formats, quantization, inference algorithms and prompt caching, were a pure-Java building block?
What if AI on the JVM could not only read and write, but also hear and see?
What if AI on the JVM were completely free, open source and built on open standards?
This talk shows how inference is built using pure Java and how to make local AI run fast on ordinary CPUs. You'll leave knowing how modern LLMs actually work under the hood, how the Vector API and quantized matrix multiplication kernels close the gap with native engines, and how to add local inference to your own Java application with a single dependency.
Come for the live demos: Gemma 4, LFM 2.5+, gpt-oss, Qwen 3.5+, Ministral 3, Granite 4.1+ and Nemotron 3+ on a vanilla JVM, at native speed. Text, audio and vision demos included.
Alfonso² Peterssen
Alfonso² Peterssen is the founder of Quixotic AI, building an AI stack from first principles for the JVM. His work challenges the belief that modern AI has to run outside the JVM.
Instead of wrapping Python runtimes or stitching together external systems, he brings AI directly into the JVM, where it can run and scale as part of the platform itself.
As a former Senior Researcher on Oracle’s GraalVM team, he created Espresso, the first spec-compliant meta-circular JVM written entirely in Java.
Today, he focuses on bringing AI and the JVM together by running the same code across CPUs and accelerators and pushing computation to the hardware that can execute it best. The approach is pragmatic: stay in Java and offload to C, CUDA, Metal, HIP, OpenCL, Mojo, and beyond when it matters. The goal is simple and uncompromising: write once, accelerate everywhere.
Through projects like Llama3.java and Gemma4.java, he demonstrated that high-performance AI implementations in pure Java are possible.
Based in Zurich, he works at the intersection of compilers, runtime systems, and AI to redefine what’s possible on the JVM.
Instead of wrapping Python runtimes or stitching together external systems, he brings AI directly into the JVM, where it can run and scale as part of the platform itself.
As a former Senior Researcher on Oracle’s GraalVM team, he created Espresso, the first spec-compliant meta-circular JVM written entirely in Java.
Today, he focuses on bringing AI and the JVM together by running the same code across CPUs and accelerators and pushing computation to the hardware that can execute it best. The approach is pragmatic: stay in Java and offload to C, CUDA, Metal, HIP, OpenCL, Mojo, and beyond when it matters. The goal is simple and uncompromising: write once, accelerate everywhere.
Through projects like Llama3.java and Gemma4.java, he demonstrated that high-performance AI implementations in pure Java are possible.
Based in Zurich, he works at the intersection of compilers, runtime systems, and AI to redefine what’s possible on the JVM.