The Number
- DeepSeek-flash now routes to DeepSeek-V4.1-Flash at $0.15/M input on cache miss off-peak, with the legacy V4-flash names retired, per DeepSeek's pricing docs on 2026-10-10.
- Qwen3.8-Max-Prime lists at 24 yuan/M input against 12 yuan for standard Qwen3.8-Max, double the token price for the premium tier, per Alibaba DashScope pricing on 2026-10-10.
- DeepSeek-flash's cheapest tier reaches $0.003/M cached input off-peak, the floor among China-lab listings, per DeepSeek's pricing docs on 2026-10-10.
Top of the Fold
- Moonshot marks Kimi K3 as formally released as its flagship on the Kimi platform, now the top tier in its pricing docs above K2.7 Code and K2.6, per Moonshot's pricing page on 2026-10-10.
- Huawei opens public beta of three HarmonyOS PC native dev tools, DevEco Studio, DevEco Code and DevEco CLI, closing the gap where a HarmonyOS PC could run apps but not build them natively, per 智东西 on 2026-10-10.
- CNCC2026 confirms its 10-21 Chengdu run with 12 academicians and 92 speakers from Tencent, Alibaba, Baidu, Xiaomi and JD, a read on who China's compute establishment is fielding this year, per 智东西 on 2026-10-10.
Untranslated
- 夕小瑶 relays that Codex lead Tibo has pledged a daily shipped improvement or a reset for 28 days, with GPT-6 Astra and 6.1 Sol sped up roughly 50% in the first four, a cadence Chinese dev circles are tracking as the pace to beat, as a named view per 夕小瑶科技说.
- A MIT-affiliated preprint circulating in Chinese AI feeds fixes the first two output tokens of Ai2's Olmo-3-7B to ".\n\nOkay" and lifts MATH-500 accuracy from 42% to 78%, above the 75% of its RL version, with no weight changes, flagged as preprint per 智东西.
- 新智元 relays that OpenAI's 2026-10-06 release of 722 math manuscripts, including a 175-page draft at the four-dimensional Kakeya conjecture, drew a joint rebuttal statement led by AMH and posted via Tao's blog, framing it as "compute hegemony," circulating/unconfirmed on the signatory list.
- SimpleMemVLA proposes giving robot vision-language-action models native visual context as memory, a China-curated preprint signal, flagged as preprint per PaperWeekly.
The Kicker
- American teenagers have reportedly turned the comment sections of NPR podcasts into a flirting ground, prompting a Chinese tech blog to note that kids elsewhere improvise group chats wherever a text box exists, per 虎嗅.
