Hi HN, Some would call AI on the JVM quixotic. And so, Quixotic AI was born.

jinfer is an inference engine for the JVM: chat, vision, audio, embeddings, reranking, and TTS. No Python runtime, no ONNX, no Docker containers, no sidecar process, no IPC. Finally, AI in a jar.

The stack underneath is built for the JVM rather than bolted onto it:

    toknroll: pure-Java tokenizers, zero dependencies
    gguf / safetensors: read and write llama.cpp and HuggingFace model formats
    jam: quantized matmul kernels, competitive with llama.cpp on CPU
    jota: Tensor API targeting Java, C, CUDA, HIP, Metal, OpenCL, and Mojo

It ships with integrations for Spring AI and LangChain4j, and is compatible with GraalVM Native Image for low-overhead, self-contained binaries with millisecond startup.

This is an early release (CPU only): I'd especially like feedback on the API surface.

Site: https://qxotic.ai Repo: https://github.com/qxoticai/qxotic