VITA-1.5

VITA-1.5 (arXiv 2501.01957, January 2025; NeurIPS 2025 Spotlight) comes from Tencent Youtu Lab with Nanjing University, XMU and CASIA, and it targets a different slice of video understanding than the captioning models: live interaction. A multi-stage training recipe progressively teaches the LLM to handle visual and speech inputs together, and the result does speech-to-speech dialogue directly — no separate speech-recognition or speech-synthesis modules — which the authors say sharply cuts end-to-end response latency. Vision capacity is preserved alongside: the paper compares against state-of-the-art models across image, video and speech benchmarks and reports strong results on all three (author-reported). The repo (VITA-MLLM/VITA) has over 2,500 stars and public weights. Note the gap the catalog cares about: the GitHub repo carries no detected license file, so commercial terms need checking before use. There is no hosted API; a demo video and eval scripts are in the repo.

• Reported competitive results across image, video and speech benchmarks vs state-of-the-art counterparts (author-reported)\n- NeurIPS 2025 Spotlight; repo has 2,500+ stars\n- No independent replication of the benchmark figures found