The Number
- DeepSeek's second round reaches 105.5B yuan (~$15.8B), with state investors at ~27% versus 2% in round one, per The Information via 智东西.
- DeepSeek-Flash cache-miss input sits at $0.15/M off-peak, half the $0.3 peak rate, per DeepSeek's pricing docs as of Oct 30.
- GLM-5.2 lists at $0.006/M prompt on OpenRouter, the cheapest China-lab input on the board as of Oct 30, per OpenRouter.
Top of the Fold
- DeepSeek's 105.5B-yuan raise draws near 30 investors with Tencent and CATL each in at 10B yuan, and chipmakers CXMT and SMIC joining as backers close ranks around the lab, per The Information via 智东西.
- Moonshot ships Kimi K3.1 Pro and sets K2-series to close to new signups Oct 31 with full shutdown planned for late December, forcing a migration, per its platform docs.
- 零一万物 relocates its headquarters to Hangzhou's Linping district on a several-hundred-million-yuan state co-investment, completing its pivot from frontier models to enterprise AI, per 智东西.
Untranslated
- A DeepSeek Harness community dev shipped dsh-Go, an Android companion that syncs DSH tasks to mobile, and crossed 1k GitHub stars in three hours, per 智东西.
- AI-for-science startup 扰动之心 (PerturbMind) exited stealth on a tens-of-millions-yuan angel round and unveiled AlphaCell 1.0, a 1B-parameter virtual-cell model scalable to 50B, company-stated per 智东西.
- An NVIDIA-MIT team released weights for Open-H3-Regenerate-2K, a video-enhancement project built on MiniMax H3, on Hugging Face, per 量子位.
- Huawei VP 袁远 told the data-storage conference that 70% of compute sits idle with data-development rate at 28.4%, arguing GPU scale does not equal effective compute, per 智东西.
The Backstory
- GeekPark argues AI Coding's real bottleneck is no longer code generation but everything after it, the task scheduling, verification loops and cross-tool handoff that still run on pre-AI plumbing, framing ByteDance's TRAE developer assistant as the next phase, per 极客公园.
The Kicker
- OpenAI's Tibo scrapped the Oct 30 Pro-200 quota cut and doubled Pro and Business Premium limits instead, declaring the company had graduated from handing out resets to actually shipping tokens, as watched by China devs per 夕小瑶科技说.
