Multimodal / unified world-model momentum (models & benchmarks)

Research, open-source, and product releases are accelerating unified multimodal models: Mistral (Leanstral, Small 4 MoE), LTX-2 (open T2V), OpenAI Sora (ChatGPT T2V), Vidu T2V updates, Nemotron 3, and compact OpenAI variants (GPT-5.4 mini/nano). Benchmarks (MM-CondChain, MMMU/MMOU, LMEB, VET-Bench, STEVO, WebVR) are pushing evaluation toward long-horizon memory, entity tracking, and cross-modal reasoning; leaderboard shifts will influence procurement and adoption.