Amap's ABot-Recon turns 12 frames into a 10,000-frame 3D scene
Two signals out of China this hour — one a new kind of spatial-intelligence model, the other proof that a frontier model can now run cheaply on domestic chips.
Alibaba's Amap has released ABot-Recon, a streaming 3D-reconstruction model that rebuilds 10,000-frame scenes while looking at only 12 frames at a time. Announced August 28 and open-sourced with weights on Hugging Face and ModelScope, the model takes a single monocular RGB video stream — no depth sensor, no pre-known camera parameters — and reconstructs the scene as it plays, "building while filming." Amap's team reports the lowest error, state-of-the-art, on the KITTI, Oxford Spires, and VBR benchmarks.
The counterintuitive trick is that ABot-Recon drops the long-range memory most streaming reconstruction depends on. Typical systems set "memory anchors" to retain and fuse history as sequences grow, which slows them down, hurts accuracy, and bloats GPU memory the longer the route gets. ABot-Recon instead solves the same bounded problem at every step, and fights drift with a lightweight motion-refiner and a pose loss designed for long horizons. The result: memory use and per-frame computation stay flat no matter how long the sequence runs.
Why it matters: autonomous driving and embodied robots increasingly operate in GPS-denied, fast-changing places — malls, warehouses, campuses — where you cannot map once and be done. A model that maps as it moves, cheaply and without the memory penalty, is the kind of primitive both industries have been waiting on. The code and evaluation harness are on GitHub under Apache 2.0, with an online demo on ModelScope.
Zhipu's GLM-5.3-Flash is now running on SenseTime's domestic compute stack. The 320-billion-parameter mixture-of-experts model — GLM-5's first natively multimodal release, with 18 billion active parameters — scored 57 on the Artificial Analysis Intelligence Index, tying Anthropic's Claude Opus 4.8. SenseTime says its heterogeneous inference approach routes different chip architectures to different stages of a request, delivering 1.25× the cost-performance of Nvidia's H-series and roughly 2.5× the token output of a same-cost homogeneous domestic setup.
The real signal here is the infrastructure, not the model itself — the benchmark debut was the story earlier this week, Z.ai's GLM-5.3-Flash tops benchmarks at one-tenth the price. SenseTime's "Token Factory" served 2.42 trillion tokens a day in July and is targeting 10 trillion daily by year-end, a 25× annual climb. If those numbers hold, the assumption that Chinese chips cannot do frontier inference cheaply is quietly dying.
What to watch: whether other model labs follow Zhipu onto domestic compute for production inference — not just demos.
Would you trust a 3D map rebuilt from 12 frames of video to guide a self-driving car? Tell us in the comments.
Sources: QbitAI — ABot-Recon · ABot-Recon (GitHub) · ABot-Recon project page · QbitAI — GLM-5.3-Flash on SenseTime
