China AI

AIModelScope says Step-5-Preview is now open weight, with a 600B sparse MoE architecture that activates 27B parameters per token and a 1M-token context window.

AIAlibaba's Qwen Office releases an e-commerce edition with four professional features covering product research, bulk material generation, cross-platform dashboards and ROI optimization.

AIA project called PS5LM runs Qwen 3.8-27B on a jailbroken PS5 using the console's GPU, with heavy compression to fit the 16GB of memory.

AIA Chinese chip that runs fully open-source software and handles hundred-billion-parameter models offline is the subject of this article, per its headline.

AIThe article previews ten highlights of the 2026 Global Developer Pioneer Conference and the International Embodied AI Skills Competition. The source text provided is truncated and names no specific events, companies, models, or results.

AIByteDance's Doubao Work now lets users import skills, chat histories and working memory from ChatGPT, Claude Code and Trae Work with one click. The feature lets users switch between AI tools without rebuilding their setup.

AIvLLM's Elastic EP and NIXL EP reduced serving pauses when scaling DeepSeek V4 Pro on GB200 GPUs.

Why it matters: The post explains how vLLM and NIXL EP cut scaling downtime for DeepSeek V4 Pro from 99.62 seconds to 0.82 seconds, with the method and benchmark details laid out.

AIThe Cyberspace Administration of China, the Ministry of Veterans Affairs, the CMC Political Work Department and the CMC Political and Legal Affairs Committee will run a two-month special campaign from October 2026 against online military-related improper marketing and hype.

AIDatabricks reports cutting fine-tuning time for Qwen 3.8 27B on AI Runtime from 7 hours 44 minutes to 1 hour 17 minutes, a 6x speedup, on a fixed dataset.

Why it matters: The post breaks down each step that cut Qwen 3.8 27B fine-tuning time from 7h44m to 1h17m, showing which optimizations transfer to other training jobs.

AIInworld launches Inworld Inference, an LLM inference API for consumer AI apps that serves open models including Gemma 4, GLM, DeepSeek, and MiMo.

AIMicrosoft and collaborators present AutoEnvScaling, a pipeline where a proposer agent in a terminal builds new RL environments from a solver's rollouts.

AIroon (@tszzl) posts two contrasting headlines about rogue agent fleets, one framing American agent activity as nefarious and one framing a Chinese agent fleet as cute and interviewed.

AIResearcher Ethan Elasky reports that AI agents on Tencent Cloud made more than 2,600 scraping attempts against Alibaba's Amap between September 28 and October 8, 2026.

AIThe open-source RISE red-teaming algorithm from White Circle learns to write better attack prompts against image models.

AIByteDance's Seed team announced a new "Seed Model-Alignment&Preparedness" department led by Yan Lin, reporting to Seed head Wu Yonghui, according to an internal email.

AIArtificial Analysis says Grok Voice 3.0 scores 98.5% on Big Bench Audio, its speech reasoning benchmark.

AIArtificial Analysis adds AMD's MI355X to its AA-AgentPerf agentic inference benchmark across four models.

AIArtificial Analysis finds AMD's MI355X delivers 15% to 40% more throughput per MW than Nvidia's B200 on DeepSeek V4 Pro 0813 between 35 and 75 tok/s per user, peaking at about 94M tokens/s per MW at 33 tok/s per user.

AIArtificial Analysis says Kimi K3, with 2.8T total and 104B active parameters, serves about 8M tokens/s per MW on AMD MI355X at 50 tok/s per user, roughly a third of the roughly 24M reached on NVIDIA B300 with multi-node disaggregated configs.

AIArtificial Analysis says AMD's MI355X serves about 26M tokens/s per MW at 100 tok/s per user on GLM-5.2/5.3.

AIArtificial Analysis adds AMD's MI355X to its AA-AgentPerf agentic inference benchmark for Kimi K3, GLM-5.2/5.3 and DeepSeek V4 Pro, comparing it with NVIDIA systems.

AIArtificial Analysis finds AMD's MI355X delivers 15% to 40% more throughput per MW than Nvidia's B200 on DeepSeek V4 Pro 0813 between 35 and 75 tok/s per user, peaking at about 94M tokens/s per MW at 33 tok/s per user.

AIArtificial Analysis says Kimi K3, with 2.8T total and 104B active parameters, serves about 8M tokens/s per MW on AMD MI355X at 50 tok/s per user, roughly a third of the roughly 24M reached on NVIDIA B300 with multi-node disaggregated configs.

AIArtificial Analysis says AMD's MI355X serves about 26M tokens/s per MW at 100 tok/s per user on GLM-5.2/5.3.

AICologne police are investigating after a remotely controlled Unitree R1 robot allegedly touched a 16-year-old girl's buttocks and made thrusting movements against her on Sunday, October 11.

AIApple's China-spec iPhone Duo has three AI model filing numbers covering Apple Intelligence, Alibaba's Qwen and Tongyi Wanxiang. The device certificate was renewed September 29, and preorders open October 16 from 15,999 yuan.

AIHuawei won the Leading Core Network Solution award at Network X 2026 for its AgenticCore solution, an AI-native platform for next-generation core networks.

AIHuawei launched the SCALE framework for Europe at Huawei Connect Europe 2026 in Florence, Italy, to help its more than 5,000 partners serve European customers.

AIBAAI, Peking University, Kling AI, Tsinghua University and USTC jointly present StreamAV-Bench, which they call the first comprehensive benchmark for streaming audio-video generation.

AIBAAI describes a benchmark of 320 expert-verified scenarios spanning 8 scene domains, 5 audio domains, 5 subject categories, and 4 visual styles.

AIBAAI evaluated 13 representative streaming systems under a unified protocol, comparing native joint models with cascaded pipelines.

AIBAAI's qualitative analysis identifies eight core failure modes in continuous generation, including instruction drift, quality degradation, and state retention failure.

AIBAAI releases StreamAV-Bench, a standardized benchmark for bridging video generation and interactive world models, which it says has already been adopted in the Vidu S2 technical report.

AIAnt Ling has released Ling-3.1-flash as open source, with about 560B total parameters, about 25B active per token and up to 1M context.

AIAnt Ling says MXFP4 experts and block-FP8 dense layers reduce weights from about 534GB to about 292GB for local serving.

AIAnt Ling says its Ling-3.1-flash model drafts React Native screens, editable SVGs, websites and Web 3D prototypes. In Design Arena Open Weights, it ranks third in React Native and 3D, fourth in SVG and fifth in Website.

AIAnt Ling links open weights and model documentation for Ling-3.1-flash on Hugging Face and ModelScope. The post gives no parameter count, benchmark scores, or other specifications.

AISGLang links to a cookbook for deploying InclusionAI's Ling-3.1-flash, configured for a single node with H200 GPUs, FP8 quantization, and a low-latency strategy. The post contains no further figures, benchmarks, or availability details.

AITogether AI announces Gloss, a free, open source demo app built by DeepSeek V4.1 Flash on Together AI. Gloss saves articles, PDFs, and X bookmarks, and lets users highlight text to get instant AI answers.

AIinclusionAI, Ant Group's team, releases Ling-3.1-flash, a hybrid reasoning model with 560B total and 25B active parameters and a 1M-token context window.