OpenAI disclosed a coordinated campaign to extract its models' hidden reasoning and linked a core group behind the activity to people associated with Moonshot AI, the maker of the Kimi language model. Researchers independently demonstrated that the same extraction technique remained exploitable on Microsoft Azure weeks after OpenAI patched its own API, exposing a protection gap across third-party cloud platforms.
Chinese-founded AI agent startup Manus has launched version 2.0 with a new personal agent app called Cue and an upgraded Studio workstation, marking its first major release since resuming independent operations after being spun off from Meta. The company confirmed it is actively building a team and preparing products for the Chinese domestic market while continuing partnerships with local model providers.
ByteDance's Doubao AI assistant has integrated flight tickets, train tickets, ride-hailing and navigation services in a full rollout announced on September 30, positioning the chatbot as a transactional life-services platform. The upgrade partners with domestic travel and mobility providers to enable end-to-end booking within the app.
A Qbitai analysis of 36 listed Chinese robot companies shows humanoid robot revenue ranging from over half of total sales to less than 0.05%, as regulators tighten IPO standards to demand verifiable revenue quality, repeat orders and cash-flow sustainability over technology demonstrations.
ESWIN Computing, a Beijing fabless semiconductor company founded by BOE creator Wang Dongsheng, is preparing a Hong Kong listing under stock code 1256 with trading expected to begin October 9. The company offers 1.57 billion H shares priced between HK$1.48 and HK$1.59 each. Despite its RISC-V computing ambitions targeting AI, automotive and robotics markets, more than four-fifths of ESWIN's 2025 revenue came from display-related chips, making the IPO a test of whether screen-industry expertise can fund a broader computing platform.
DeepSeek has open-sourced infrastructure components optimized for Huawei's Ascend 950 AI chip platform, including a compiler toolchain, high-performance operator libraries and a distributed communication library. Benchmark results for DeepSeek-V4.1-Flash on an Ascend SuperPoD cluster show per-card output throughput of 2,469 tokens per second at 5ms time-per-output-token in offline inference mode.
