AI Lab
Official Tencent Hunyuan announcements — model releases, product launches and research, each one summarized with every source covering it, by AIBriefs. RSS
Hy-SimulMT streams speech continuously with roughly three seconds of translation lag on the RealSI long-form audio benchmark, scoring above other open-source models and four of five commercial APIs compared. Release includes weights, inference software and training tools; speech output and voice cloning need a configured TTS service.
Hy-Nomas is a video world model that generates a world a few seconds at a time, caching earlier segments' keys/values so it follows multi-turn text and camera instructions while preserving interaction history. It runs natively in ComfyUI on just 12GB VRAM.
Tencent's Hy-Memory-30B-A3B is a 30B-parameter mixture-of-experts model with 3B active parameters, posted to Hugging Face. A Reddit commenter highlights its memory focus, claiming the 3B-active model beats Opus 5 at memory.
The 5B model takes a single input — a document page, natural image, chart, flowchart, geometry figure, audio clip, or audio-visual video — and returns one structured JSON envelope covering perception and layout elements.