Baidu Apollo released ApolloWorld, billed as the first Agent-driven omni-modal world model for physical AI, moving urban road networks, multimodal sensors and traffic博弈 into the cloud with millisecond-level interaction and tens-of-minutes continuous dispatch simulation, breaking past world models limited to short isolated clips. Baidu claims cloud-based virtual mileage of 100,000 km per day and that its L4 fully driverless end-to-end model is now deployed in vehicles.
Omni-modal
全模态
Model architecture capable of simultaneously processing diverse inputs like text, image, and audio, and generating corresponding outputs.
Curated Intelligence
VeRL-Omni v0.2.0 has been released, focusing on improving the throughput of diffusion model reinforcement learning (Diffusion RL) and the stability of omni-modal training. The update significantly accelerates the rollout process for models like Qwen-Image FlowGRPO by integrating vLLM-Omni and the verl V1 trainer, while introducing reusable model adapters and FSDP2 support to enhance system robustness.
Alibaba's Qwen team released Qwen3.5-Omni, a new-generation native omni-modal large model featuring a Hybrid-Attention MoE architecture for unified understanding and generation of text, image, audio, and video. The model supports 256k context length, handles over 10 hours of audio and 400 seconds of 720P video input, and significantly enhances multilingual speech recognition (113 languages) and synthesis (36 languages). The Plus variant achieved State-of-the-Art results in 215 sub-tasks/benchmarks, marking a major leap from stitched multimodal systems to a truly native unified architecture.
Alibaba's Qwen team released the updated Qwen3-Omni-Flash-2025-12-01, significantly improving audio-video instruction understanding to resolve performance degradation in colloquial scenarios. The new version fully opens up System Prompt customization for precise control over personas and styles, while enhancing multilingual adherence stability across 119 text languages and more natural speech generation. This native omni-modal model is now open-sourced on platforms like Hugging Face and ModelScope, supporting real-time streaming text and voice outputs.