Research addressing RL training of LLMs without requiring ground-truth solutions. Addresses cost of RLHF annotation.
audio.cpp consolidates 12 audio models (Qwen3-TTS, PocketTTS, VeVo2) in single C++/ggml runtime with 5x faster TTS inference.
Outerport enables dynamic model weight swapping without service interruption. YC S24 company with 93 HN points.
JetSpec enables parallel tree drafting in speculative decoding achieving up to 9.64x lossless LLM inference speedup with 1000+ TPS throughput.