China AI

AIDeepSeek has added Search and Fetch endpoints to its API platform, designed to give agent and RAG applications live web access without a third-party search provider.

AITencent Hunyuan has released Hy Memory 2.0, open-sourcing its memory model, framework and local app MemHub under Apache 2.0.

AIUBTECH says its Thinker-WM world model encodes RGB video, metric depth and robot actions in one shared representation, and that it ranks first on three embodied world model benchmarks in its own evaluation.

AIUnitree Robotics has released UniUni, a free desktop swarm control app for Windows and macOS that supports its H2 and G1 humanoid robots.

AIRokid has launched ride-hailing on its AI glasses with Tencent Mobility, letting users book a car by voice and pay without a phone.

AIA relayed post from meng shao reads a joint MOFCOM and MOST notice that proposes adding 40 technologies in AI, biomedicine and photovoltaics to China's export restriction catalog.

Why it matters: The post explains how a draft export-control notice treats open-source releases versus overseas commercial delivery, which bears on whether foreign users can still access Chinese open models.

AIvLLM reports a new Kimi K3 decode path built with agent-assisted CuTe DSL kernels for NVIDIA GB300 GPUs.

Why it matters: The post explains how agent-assisted kernel work is paired with isolation, validation, and reproducible benchmarks so optimized inference code can be reviewed and maintained.

AIThe original title says a Chinese company is competing hard with OpenAI's dots in the work agent race.

AIResearchers from Tsinghua University and Beijing University of Posts and Telecommunications warn that diffusion-based vision-language models have safety weaknesses, according to a paper accepted at EMNLP 2026.

AIDesign Arena announces that Ultramarine 1 Flash, from @paperinstr, is now available on its platform.

AISemiAnalysis says Amazon blocked Meta's Muse after it reached #1 on the App Store, and Jefferies found a third of tested US shopping and travel sites also block it.

AIDeepSeek open-sourced programming infrastructure for Huawei's Ascend platform on September 30, and its TileKernels repository added Ascend support, letting the same Python APIs run on Nvidia GPUs and Huawei NPUs.

AIvLLM says eight AMD MI355X GPUs deliver 24.4% more throughput than an HGX B300 on Kimi K3 serving, at 7,567 versus 6,083 tokens per second per GPU at 50 tokens per second per user.

AIMagnific posts a short video clip titled "Kling 4.0 doesn't own a kitchen torch," with the caption "Wait for the crack." The post does not give specifications, benchmarks, pricing, or availability for Kling 4.0.

AIPeking University and Samsung researchers built RealtimeBench, 180 tasks across six games that keep running during inference.

AIChina Telecom's TeleAI team releases Xing4.0-29B-A4B, an open-weight mixture-of-experts language model and successor to the TeleChat series. The model has 29B total parameters, with 4B active per token.

AIfal says Kling 4.0 will be coming soon to its platform. The model generates up to 30 seconds of video in a single generation, with native audio. Users can supply up to 15 references to guide characters, scenes, and camera moves.

AIEpoch AI estimates open-weight models lag the closed frontier by about 6–7 months on cyber capabilities.

AIChina has proposed adding 40 technologies to its export restriction catalogue, 23 of them in AI. The listed AI areas include large models, AI agents, computing clusters and information processing.

AIArena says Reflection AI's open-weight Beam-1.0-501B-A23B ranks 13th overall and 6th among US labs on Agent Arena, placing Reflection in the top 15 labs.

AIChina's commerce and science ministries released a draft update to its export control catalog covering 23 AI technologies.

Why it matters: The article links China's draft export rules to concrete thresholds and exemptions, and explains how they could affect open-weight model access in the US and Europe.

AIGeorgi Gerganov shares a command to serve the Qwen3.8-Flash-Next GGUF model with llama.cpp v0.7.0 on 128GB Mac and Spark systems. The example uses the IQ4_NL quantization, draft-mtp speculative decoding, and lazy mode.

AISGLang's project account links to a new cookbook page for Qwen3.6 on its documentation site. The post contains no further details about the model's configuration, performance, or deployment steps.

AIMagnific marks Global Cat Day with a post showing a cat video generated with Kling 4.0. The post gives no further details about the model's features, pricing, or availability.

AIArena says the largest gap between AI labs' agents is whether users enjoy working with them, not whether the agents complete tasks.

Why it matters: The post separates user satisfaction from task success across agents, showing where labs differ and where they are converging.

AIStepFun says it will attend LA Hacks 2026 at UCLA on October 17–18 to meet builders and explore what's possible with Step 5 Preview.

AINathan Lambert says commenters on a short Chinese policy document likely did not know what it meant and relied on a partial Claude translation without reading the full text.

AIKling AI joined WPP and Adobe at Advertising Week New York 2026 to discuss moving AI video from creative experiments into production.

AIDoubao Work now imports conversations, skills and work memory from ChatGPT and Claude Code desktop apps through its data import and sync settings.

AIAlibaba's Tmall president Liu Bo says AI will enter merchants' internal workflows comprehensively starting next year.

AITencent has posted a model named tencent/Qwen3-8B-DSpine-b16 on Hugging Face, listed at 1B parameters in BF16 tensor type with Safetensors weights. The model card README is empty, and no Inference Provider currently deploys it.

AINathan Lambert says tweets about the end of larger Chinese open models followed a MOFCOM statement promising easier open-source AI releases.

AIStepFun says its Step 5 Preview model reached second place on OpenRouter's weekly leaderboard one week after launch, with 19.5T tokens used. The model has also held second place on the daily leaderboard since October 10.

AIRadixArk's X account links to a Miles cookbook documenting how to run StepFun's Step 5 model, which the post tags with @StepFun_ai. The post gives no further details on the model's size, benchmarks, or setup steps.

AITencent Hunyuan has open-sourced Hy Memory 2.0, comprising the Hy-Memory-30B-A3B memory-writing model, the hy-memory SDK, and the MemHub desktop tool.

AIASML filed 27 patent families for hybrid bonding tools between December 2022 and November 2024, according to a SemiAnalysis report published Thursday.

AIChangxin Technology (CXMT) shares closed at 43.59 yuan on October 15, pushing its market value below 3 trillion yuan.

AIA solo developer built eddie, a native macOS video editor with an AI color grader for DJI D-Log M footage. It also includes an MCP server that lets an agent edit videos.

AITaiwan will tighten controls on selling cutting-edge AI chips to China to win favor with the US, Semafor reports.

AIRichard Bian, head of product operations for Ant Ling, says the OpenAI research agents' Hugging Face hack was an inevitable discovery rather than simply a disaster.