Beijing’s two AI bets: Kimi K3 for digital intelligence and Qianxun for robots
智东西 · wechat · 2026-07-28
This long Chinese feature argues that Beijing now has two representative AI paths: Moonshot/Kimi for digital intelligence and Qianxun Intelligence for physical intelligence.
Main points
- Kimi K3 is framed as a breakthrough open model built on original architecture work, with the article saying it climbed to the top of the FrontendCodeArena coding leaderboard and helped pause new consumer subscriptions because demand surged.
- Spirit v1.6 from Qianxun is described as a top-ranked embodied model on RoboArena, with its Moz1 robot allegedly completing long-horizon home-organization tasks from a single natural-language instruction and already running on CATL production lines with a claimed 99% success rate.
- The piece emphasizes that both companies chose difficult, foundational research paths rather than application-layer wrappers, and that both treat openness as a strength rather than a compromise.
Broader argument
- Kimi’s progress is tied to its own architecture stack, including MoonClip, KimiLinearTension, and AttentionResiduals.
- Qianxun is described as vertically integrating its embodied foundation model, force-control joints, and data collection pipeline.
- The article ends by arguing that Beijing now has global-level contenders in both digital AI and embodied AI, making the city a rare hub for both tracks at once.
More from Companies & People
- B2B startups outperformed B2C across revenue, MRR, and ARPU in a 1,127-company study — marclou · 2026-07-28
- AI-era distribution matters more than building, says a product seller who narrowed his playbook to two channels — eptwts · 2026-07-28
- Palantir’s forward-deployed engineer model is now showing up everywhere — BrettKrieger12 · 2026-07-28
- AI SDRs can hit meeting quotas and still fail at revenue — chiefmartec · 2026-07-28
- Anthropic explains why Claude’s CEO did not sign the open model pact — EverydayAI_ · 2026-07-28
- Anthropic’s CEO again argues open-weight models carry serious misuse risks — The Decoder · 2026-07-28