The Overnight

The Overnight — 2026-08-15

Qwen3.8-27B goes open on consumer GPUs; GLM-5.3 and DeepSeek Harness crowd the coding lane

What China's AI world did while the West was asleep. · 2026-08-15

Share: Share on X LinkedIn

The Number

  • DeepSeek-V4-Pro-0813 cache-miss input runs $0.435/M against $0.87/M output, before a peak/off-peak split takes effect 16:00 UTC Aug 16, per DeepSeek's pricing docs (as of 2026-08-15).
  • Qwen3.8-27B carries 27B parameters, small enough to run quantized on a consumer GPU under Apache 2.0, per Alibaba (2026-08-14).
  • Qwen models have crossed 3B cumulative downloads worldwide across 460-plus open releases, per 雷峰网 citing Alibaba (2026-08-14).
  • On OpenRouter, qwen3.7-flash lists at $0.03/M prompt tokens, the cheapest China-lab entry on the board (on OpenRouter, as of 2026-08-15).

Top of the Fold

  • Alibaba open-sources Qwen3.8-27B, a native multimodal dense model that quantizes onto home GPUs, lowering the local-deploy bar below its own MoE flagships, per 量子位 and 智东西 (2026-08-14).
  • Zhipu ships GLM-5.3, a 743B coding-focused model it says trades blows with Kimi K3 and pledges to open within two weeks, per 智东西 (2026-08-14); vendor benchmark claims are company-stated.
  • DeepSeek opens DeepSeek-V4-Pro-0813 alongside its MIT-licensed Harness agent framework, its first agent product, framing model-plus-harness as the unit that equals an agent, per 极客公园 (2026-08-13).
  • Reuters reports Apple is co-training a China-market model with Alibaba, days after an Apple Mac manual quietly documented then deleted a Qwen integration, per 智东西 (2026-08-14).

Untranslated

  • DeepSeek Harness clears 50,000 GitHub stars within 12 hours of the repo going public, with 极客公园 testing it to rebuild its own site and 虎嗅 arguing the real bet is whether an AI can rewrite itself, per 极客公园 and 虎嗅 (2026-08-14).
  • Meta researcher 余家辉, one of seven Chinese hires named to its superintelligence lab, quits after one year to start his own venture on an undisclosed direction, per 量子位 and 智东西 (2026-08-14); circulating on X, unconfirmed on specifics.
  • Baidu's Wenku-Wangpan agent GenFlow gets a Chinese name, 库库AI, and a standalone office client, claiming 25M+ MAU on AI office work, per 雷峰网 (2026-08-14); company-stated.
  • CUHK opens Libra, an agentic-RL resource manager claiming up to 300% throughput and 2.5x faster training, per PaperWeekly (2026-08-14); preprint claims, unverified.
  • SenseTime's compute arm coins a TPW metric, pitching data centers as a link between electricity and tokens rather than a PUE box, per 雷峰网 (2026-08-14); one vendor's framing, unverified.

The Backstory

  • Insta360, the panoramic-camera maker that watched buyers leave the gadget in a drawer after a few ski trips, is pushing its X6 toward parents and road-trippers so the camera disappears into the trip, per 极客公园 (2026-08-14).

The Kicker

  • Alibaba held its smart-driving forum inside the Ulanqab data center that runs 60% of China's autonomous-driving training, telling carmakers the model on the road is running under their feet, per 极客公园 (2026-08-14).