AI sovereignty for the world's most trusted runtime.
The JVM powers global finance, big data platforms, and mission-critical enterprise infrastructure.
It deserves an AI stack built to the same standard.
We're building Jota to be the JVM's missing AI substrate: building blocks that let Java run AI at native-speed without sacrificing the portability and reliability the platform is famous for.
qxotic is a modular AI stack, it provides the following standalone components:
| Module | Purpose |
|---|---|
| jota | Multi-backend tensor engine (Panama, C, CUDA, HIP, Metal, OpenCL, Mojo) |
| gguf | Native read/write for llama.cpp's GGUF model format |
| safetensors | Native read/write for HuggingFace's Safetensors format |
| Tok'n'Roll (toknroll) | TikToken-compatible + customizable BPE tokenizers for popular LLM models |
Jota provides the computational foundation, abstracts hardware complexity behind a unified Tensor API. Write once, accelerate everywhere, from embedded devices to data center GPUs, without leaving the safety of the JVM.
First-class support for GraalVM's Native Image. On JVM, Jota's host backend defaults to Panama (JIT-compiled Java), on Native Image it uses the C backend instead. Except for Panama, all backends are supported out-of-the-box on Native Image with minimal overhead e.g. a simple Mandelbrot demo with ALL backends included is just ~22MB.
Quixotic AI is the evolution/realization of llama4j a complete set of standalone JVM components for AI.
Experimental. Evolving. Under heavy development.
- 🌐 qxotic.ai
- 🐦 @qxoticai
- 🦋 @qxotic.ai
Built in Switzerland. Accelerating (on) the JVM.