Pure Java 21+ CPU inference (continuous batching, paged KV cache, safetensors / GGUF / ONNX): chat, BERT embeddings, Whisper speech-to-text, and Piper text-to-speech. Inspired by GeeeekExplorer/nano-vllm