VLA Decoding Notes (4): Domestic Players — Ant Lingbo, Xiaomi, Tencent

0. Clarify a Mix-Up First

The “Xiaomi also has Hy3” you mentioned needs correcting: Hy3 / HyVLA belongs to Tencent’s Hunyuan embodied team, not Xiaomi. Xiaomi’s VLA model is Xiaomi-Robotics-0. We cover all three separately below.

1. Ant Lingbo: LingBot-VLA 2.0 “One Brain, Many Machines”

• Background: Ant Lingbo Tech, a wholly-owned Ant Group subsidiary, founded 2025-03 in Shanghai Pudong, focused on the robot “general brain”. Started its first independent funding round in 2026-08, targeting ~¥1.5B.
• Flagship LingBot-VLA 2.0 (released 2026-07): “one brain, many machines” — pre-trained to adapt to 17 brands (Unitree, AgiBot, Galbot, Leju, etc.), 20+ robot configurations, showing cross-embodiment reuse and generalization.
• LingBot model matrix: VLA 2.0 (base / cross-embodiment), VA 2.0 (world-action model), Vision (spatial-native visual base model), World 2.0 (infinite-duration real-time interactive world model), Video (MoE open video generation).
• Industry take (CEO Zhu Xing): in past years “cerebellum” and hardware advanced fast, but the brain is the bottleneck; future “AI will redefine hardware”.

2. Xiaomi: Xiaomi-Robotics-0 (Open Real-Time VLA)

• Released: open-sourced 2026-02, 4.7B params, announced by Lei Jun, weights/HF/code all public.
• Architecture: MoT — VLM brain + 16-layer DiT cerebellum, KV Cache loosely coupled (see Part 2).
• Performance: 80ms latency, 30Hz, real-time on RTX 4090, SOTA on LIBERO / CALVIN / SimplerEnv.
• Highlight: solves “action discontinuity” via async inference + Action Proposal + Clean Action Prefix + Λ-shape attention.
• Significance: a big lab open-sourcing a consumer-grade real-time VLA lowers the industry barrier and feeds a factory data flywheel (90.2% → 98% frame-by-frame optimization).

3. Tencent: HyVLA and FlowPRO

• FlowPRO: uses flow-matching regression loss as an implicit reward proxy, no extra Critic network; four long-horizon bimanual tasks hit 94%–99% success, beating traditional DAgger’s 83%–93%.
• Data: direct optical motion-capture, sub-millimeter precision.
• Positioning: Tencent Hunyuan (Hy family) methodology output for embodied AI — emphasizing that post-training innovation affects deployment more than parameter scale.

4. Landscape

Common trend: the “compete on brains” stage — whoever first has a complete model system + real-scene data loop runs out a platform-level company. Ant Lingbo already plugs 17 brands into one brain, an upstream “brain supplier” logic over cerebellum/hardware vendors (Unitree, etc.).

5. Bridge

Domestic players push VLA toward “cross-embodiment + open-source + engineering loop”. Next: VLA’s next stop — world models, Sim-to-Real, data flywheels, and the scaling bottleneck.