The Overnight

The Overnight — 2026-09-11

DeepSeek ships V4.1 Flash, the KV-cache killer

What China's AI world did while the West was asleep. · 2026-09-11

Share: Share on X LinkedIn

The Number

  • DeepSeek V4.1 Flash cache-miss input runs $0.15/M off-peak ($0.30 peak), against V4 Pro's $0.66, per DeepSeek's pricing docs as of 2026-09-10.
  • Qwen3.7-flash sits at $0.03/M prompt on OpenRouter, the cheapest China-lab listing there as of 2026-09-10.
  • Ant's 阿福 health app reports 150M users and 20M daily consultations, per 百灵 at the Inclusion Bund summit on 2026-09-10.

Top of the Fold

  • DeepSeek releases V4.1 Flash, a 552B-parameter MoE on a new architecture that it says surpasses V4 Pro while cutting price, with V4 Pro to be retired into it from 2026-09-14, per DeepSeek's docs and 智东西.
  • Alibaba's Amap ships ABot-Earth 0.7, a 3D-native city world model that generates kilometer-scale scenes in ~10 minutes on one consumer GPU, its bid to move mapping from language to spatial intelligence, per 智东西.
  • WeRide wins Spain's first L4 robotaxi passenger license with Uber and AVOMO, its opening into commercial European operations this year, per 雷峰网.

Untranslated

  • Ant's 灵波 open-sources a 1.3B world model that runs real-time on a single consumer GPU, letting two users co-drive one generated scene as player and director, per 夕小瑶科技说.
  • JD open-sources JoyAI-Echo WM, an interactive world model paired with AI glasses feeding first-person scenes to a VL model for blind users, backed by a stated 100,000-card domestic compute cluster (company-stated), per 量子位.
  • WeChat quietly tests 小微 AI social, where two users' agents negotiate a request in a separate space and return only for confirmations, per GeekPark's own hands-on test (attributed as own-brand).
  • 新智元 flags a circulating claim that Terence Tao posted anger that AI is killing math's century-old open tradition, with mathematicians afraid to disclose directions; circulating/unconfirmed.

The Backstory

  • Hypershell founder 孙宽 tells GeekPark the exoskeleton maker shipped over 30,000 units globally in 2025 and is now pushing an early, niche category toward a mass consumer product, per 极客公园 (attributed as own-brand).

The Kicker

  • 夕小瑶's take on the 灵波 world model: walk down a corridor and your friend spawns the monster, tire of sunshine and just ask it to snow, the scene regenerating as you go, per 夕小瑶科技说.