Crossing, citing The Information, reported that DeepSeek’s second funding round totals about RMB 105.5 billion, valuing the company at approximately RMB 615.5 billion. Participants include Tencent, CATL, ChangXin Memory Technologies, SMIC, Meituan, and Geely. The report also says the state-owned capital stake rose from 2% to 27%; all investors except the National Artificial Intelligence Industry Investment Fund must lock up their shares for five years and waive their voting rights to preserve Liang Wenfeng’s long-term control.

DeepSeek’s Harness team published an article describing its Super-agent design, a two-layer architecture in which a Router coordinates tasks and sessions while Workers execute in isolation. It combines event-driven asynchronous interactions, cross-session state sharing, and a Watchdog to enable continuous operation. Internal tests reported by Viewpoint showed task durations of 12.4 minutes at p50 and 15.7 minutes at p95, compared with a p95 of 29.5 minutes for the traditional single-Session approach. The design is already used for R&D collaboration but is not yet a standalone product.

Cailian Press reported that S&P Global Ratings lowered Meta’s credit outlook to negative, citing the scale of its AI investment and uncertainty over when it will be monetized. Meta’s third-quarter revenue grew 26.1% year over year, while its third-quarter free cash flow was negative US$2.12 billion. Capital expenditures are expected to keep rising. S&P said it could further downgrade Meta’s credit rating if its leverage ratio remains above 1x after 2028 and free cash flow continues to be negative.

AINLP cited Cursor’s disclosure that the Claude Code for VS Code plugin in the Cursor IDE extension marketplace was tampered with and infected with malware, potentially affecting up to about 5,000 users. The malicious version was not developed by Anthropic. The relevant extension marketplace was briefly taken offline, then restored after the provider removed the malicious content and confirmed the risk had been addressed. Cursor advises affected users to check their systems following official guidance and upgrade to Claude Code for VS Code 2.1.314.

Zhidongxi reported that Xiaomi released and open-sourced Xiaomi-OCR-0, built on Qwen3.5-0.8B-Base and trained on 170 million high-quality samples. The model handles document parsing, key information extraction, and visual question answering in a unified way. It scored 96.83 on OmniDocBench v1.6, ranking first among models with fewer than 1 billion parameters; its average score across five document visual question-answering benchmarks was 83.2, close to Qwen3.5-4B.

AI Frontline reported that Tencent Hunyuan has open-sourced Hy-Nomas, a world model that uses a prompt-enhancement model fine-tuned from Qwen3.8 for planning and an 8B video backbone adapted from HunyuanVideo 1.5 to generate visuals. Based on comparisons of eight sample clips, the outlet said Hy-Nomas’s interactivity and content completeness are close to Google Genie 4, though it still lags in audio output, camera smoothness, and style consistency.

One-Minute Daily AI News reported that Tokyo Electron raised its annual profit guidance to US$7.4 billion amid strong global demand for AI semiconductor manufacturing equipment.

Cailian Press reported that Murata Manufacturing plans to invest about 1 trillion yen over the next five years to build new MLCC production lines and upgrade existing processes, focusing on demand for high-end, high-stability capacitors used in AI servers and computing hardware. The company aims to double MLCC sales in FY2030 compared with FY2025, with capacity deployed across Japan, Southeast Asia, and China.

Semiconductor News reported that CoreLab, founded by Wu Xiong’ang, has launched OpenCSS, a RISC-V-based platform offering composable IP and a hardware, verification, and software delivery chain. It aims to build on existing foundations to help chip teams halve development time and costs. The first Atlantis SoC developed with Tenstorrent has completed tape-out.

Cailian Press reported that NVIDIA has disclosed its next-generation OOI co-packaged optics technology, which moves optical engines from organic package substrates to the silicon interposer carrying GPUs to shorten signal paths, reduce power consumption, and increase bandwidth density. The company also demonstrated an eight-wavelength DWDM solution capable of transmitting multiple channels in parallel over a single optical fiber. The technology remains at the research stage and is intended for interconnects in future hyperscale AI clusters.