Open models surge (text, image, video, audio)

New open-source KVAE tokenizer family for audio, image, and video outperforms FLUX.2, HunyuanVideo; code released. Earlier: DeepSeek Flash v4, JoyAI-Video-Edit, Hunyuan3D-Buffalo 1.0, Qwen3.8-Max, NVIDIA Nemotron, MiniMax H3, DeepSeek V4, Kimi K3. Open-weight ecosystem continues to expand.