WeituAI's streaming video brain scores 69.9 on OVBench, beating Gemini baseline by 11 points
机器之心 · wechat · 2026-11-16
Hong Kong multimodal startup WeituAI introduced a streaming video brain that gives camera-equipped devices continuous perception and memory. It serves dozens of enterprises and 6M individual users, the team says.
Key points:
- DualSparseMemory architecture: temporal block-level sparse selection first narrows attention to relevant history segments, then spatial sparse learning extracts key visual details, controlling compute and cache overhead for long-running devices.
- Long-term Memory Module compresses older spatiotemporal information into compact, continuously updated representations, supporting cross-time reasoning ("where did I put the keys") and even prediction of unfinished action boundaries.
- Benchmarks: 69.9 composite on OVBench (16 subtasks spanning past memory, current perception, future prediction), 11.1 points above the stated Gemini 3.1 Pro baseline (58.8); 58.23 on egocentric-memory Meta S-EMBER vs 54.41 for the aligned baseline.
- Use cases include AI glasses, industrial maintenance assistance, video creation, and acting as a continuous perception front-end for general LLMs and future robots.
More from Embodied
- Rabbit R1 is a weird device in all the best ways, says dev — SimonBalmain · 2026-11-16
- Slow models are a disaster in robotics: researcher dives into fixing latency for large VLAs — kvablack · 2026-11-16
- miHoYo unveils Gamer, a VLA foundation model for real-time game agents — Run Luo · 2026-11-16
- Grid Robotics Proposes Agentic Robotics Paradigm Built on Embodied System 2 — Grid Robotics · 2026-11-16
- Shenzhen hosts APEC 2026, showcasing humanoid robots and its design-to-mass-production cycle — CyberRobooo · 2026-11-16
- Tongxin Intelligent raises hundreds of millions RMB to build AGI-native inference chips — 量子位 · 2026-11-16