China AI Dynamics

Ant Group's Ling-3.1-flash has about 560B total parameters with ~25B activated per token and a 1M context window, continuing the hybrid-linear architecture and raising the linear-attention ratio (7 KD

DeepSeek open-sourced infrastructure components for Huawei's Ascend compute platform, including the TileLang compiler, DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect, mirroring its earlier NV

The author summarizes OpenAI DevDay 2026: the personal agent product Dots launched to ChatGPT Pro, Business Premium and Enterprise users with 4,000+ app integrations; new GPT-6.1 Sol went live at abou

China's first AI long-form drama 'Journey to the West: Aftermath' landed on Hunan TV prime time on Aug 31; planned at 60 episodes of ~40 minutes each, shot with no cameras, 100% video generation by Se

Xiaomi released MiMo-V2.6 Pro and Flash, open-weight full-modal models trained via large-scale trial-and-error learning; Pro scores 46 on the Artificial Analysis Intelligence Index, the highest among

Ant Group released the open-source Ming-Image-0.1-Design series with two 6B models: Design generates complete designs such as UI, infographics, and posters from text, while Layer decomposes a design i

Chinese AI lab StepFun released Step 5 Preview, a flagship base model using a sparse multi-expert architecture with 600B total and 27B active parameters, supporting a 1-million-token context and text

A reported incident describes US forces almost acting on a fabricated intelligence assessment produced by an AI system, highlighting the danger of hallucination in high-stakes military decisions.

TypeSafe AI released Jev, a specialized model for high-frequency decisioning; a hands-on test shows strong cost-performance on classification judgment.

An independent author shared a workflow that saves Codex quota: wrap your own server as a read-only, least-privilege, Feishu-OAuth MCP Server that GPT-6 Pro on the ChatGPT web app calls as a plugin, r

Trade data matches roughly three billion dollars of chips smuggled into China via Malaysia, pointing to a gap between official figures and actual flows.

Shengshu formally launched Vidu S2, including Vidu S2-Avatar for real-time interaction with digital characters and Vidu S2-Editing for real-time editing of video streams, and is exploring real-time sp

StepFun released the StepAudio 3 family with Realtime, ASR, TTS, Gen, and Music models, live on the StepFun open platform, with several variants ranking first globally on Artificial Analysis leaderboa

Xiaohongshu's AllSpark team released the open-weight Search Agent model Iris; weights and eval code are public, with data and training recipe to follow. The 35B and 397B versions lead at their scale.

Anthropic released report 2025 12 2026 8 Claude spyware organizations AI organizations Claude Code 2000 missile FPV drone

DeepSeek V4.1 Flash cache 7 9 14 12 v4-pro 4.1 Flash

Anthropic released report alleged AI Claude model model 2

27 research Jacob Coxon said Open AI Anthropic self-improvement Anthropic AI supports Tim Urban 2015 The AI Revolution extinction

NSA FBI CISA alleged DeepSeek Moonshot AI Moonshot AI Alibaba MiniMax StepFun Z.ai 2024 model extraction said coding capability model model

NSA CISA FBI released AA26-251A alleged DeepSeek Moonshot AI Alibaba MiniMax Z. AI 2024 model model model

Open AI shipped GPT Image 2.5 generation editing 15 image call interface released GPT-Image-2.5 Flare 50% GPT-Image-2.5 Sunburst

GPT-6 Astra inference Reasoning Effort model Ultra AI collaboration

released GPT-6 Astra Blender MCP Computer Use Computer Use 4 200 Pro MCP CLI Python

GPT-6 Astra shipped users Blender Houdini Unity Aseprite Demo 3D Astra

On September 2 UU Remote shipped an update focusing on terminal capabilities: full TUI rendering and multi-terminal session management. Highlights include passwordless Mac login, mobile input optimiza

MiniMax connected its H3 Max model at 768P and 480P to its open platform and MiniMax Design. Overseas developers have already used the APIs to build a Twitch livestream and a 24-hour 'AI TV station' t

Zhipu AI has open-sourced the weights of GLM-5.3, which supports local deployment and customization and excels at complex coding, defensive cybersecurity, and long-horizon tasks. It scored 60 on the A

Tencent Hunyuan has released its new flagship model Hy4 preview, with 770B total parameters, 49B active parameters, and a 1M token context length. The model is open-source and freely usable, and is no

As of June 2026, China's daily token volume passed 500 trillion, with its large models now in the global front rank.

MiniMax open-sourced its H3 video generation model. The LMSYS joint team, with NVIDIA and Ant Group, benchmarked it on 8x H200 GPUs, achieving up to 1.95x lossless speedup and a peak 6.24x with SSIM b

Doubao Work is the lowest-barrier path for enterprises to adopt agents but needs a Feishu login for full features; it can remotely control up to seven devices, run scheduled tasks, read local skills,

Tencent Hunyuan compressed its on-device translation model Hy-MT2-1.8B to 574MB and 440MB via 2-bit and 1.25-bit quantization with near-lossless quality, beating Microsoft Translator on FLORES-200, ad

ByteDance released Doubao Work, the first major product after Feishu and Doubao teams merged, natively integrated with Feishu and able to encapsulate routines as reusable Skills.

Zhipu open-sourced GLM-5.3-Flash (320B-A18B), the first natively multimodal GLM-5 model, scoring 57 on the AA index on par with Claude Opus 4.8, at one tenth GLM-5.3 and one fortieth Opus 4.8 during p

ModelBest's OpenBMB team has released MathForm, an open-source framework, dataset, and model suite for automated mathematical formalization in Lean 4. Its FormalVerse dataset contains over 367K verifi

Alibaba has officially launched Qwen-UI Agent, a real-world-centric GUI agent foundation model covering mobile, desktop, web, and DeepSearch environments.

GLM-5.3 API is live today, excelling at complex coding, defensive cybersecurity, and long-horizon tasks. It scores 60 on the Artificial Analysis Intelligence Index, on par with undisclosed internal fl

Inspired by a Reddit post about making Claude truly start thinking, the author introduces the 'steelman' concept from logic and devises a 'bidirectional steelman prompt for AI.' Through four steps res

Qwen has open-sourced the Qwen3.8 model series. Qwen3.8-27B is a native multimodal (image/voice/text) dense model whose 27B parameters already reach the level of Qwen3.7-Plus, with native 262K context

Xiaohongshu Technology has open-sourced dots3-note Preview, the lightest model in the dots3 family. With 280B total parameters and 16B activated, it supports 512K context and understands text, vision,

Zhipu has released GLM-5.3, built on the same base as GLM-5.2 but raising the intelligence ceiling through extreme post-training scaling. Coding improved 50% over the previous generation, ranking firs

Ant Lingma and the ASystem team collaborated to run a complete single-machine agentic RL post-training loop on a DGX Spark, using Ling-3.0-tiny and AReno. With tic-tac-toe as the minimal validation ta

From January to August 2026, public model repositories on Hugging Face grew from 2.43 million to 2.96 million, yet 85.6% of models were downloaded fewer than 200 times, and 1.5% of repos captured 99.2

A beginner-friendly walkthrough of the DeepSeek Harness, an agent framework that orchestrates models, tools, skills, and loops into reusable workflows. The piece covers what the Harness is, its core c

Xiaohongshu (RED) dots team open-sourced dots.tts, a 2-billion-parameter fully continuous end-to-end autoregressive speech synthesis model, achieving the best average content accuracy and average spea

WorkBuddy updated with a remote control feature that connects PC, app, and mini-program, letting the phone sync in real time the tasks, conversations, workspace, and artifacts from the computer, suppo

DeepSeek V4 Pro and xAI's Grok 4.6 were released within a two-hour window, with roughly 1.6T and 1.5T parameters respectively, both nearing the coding experience of Claude Code.

MiniMax launched Music 3.0, a new-generation music generation model that can complete the composition, arrangement, performance, and production of an entire song at once from a creative concept and op

Alibaba's Qwen team open-sourced the Qwen3.8-2.4T-A95B model weights, its first Qwen-Max-level model released for free use, with 2.4T parameters, 95B active per token, and a context expandable beyond

The article lays out a 12-step practical workflow for absolute beginners to get started with AI in half a day: prepare a computer with at least 16GB of RAM, subscribe to ChatGPT and install Codex or u

With new cross-session messaging in Codex and Claude, a main-plus-branch conversation structure replaces handoff docs and git backups, reshaping how developers collaborate with AI on code.

Ant Ling (Bailing) has open-sourced Ling-3.0-tiny, a native hybrid reasoning model with 7.9B total parameters that activates only 1.3B parameters during inference, and simultaneously releases three ve

ZCode, deeply optimized for GLM, launched four features today: Goal, Subagents, Remote Control and idle tasks. In the Z.ai Code Bench test, GLM-5.2 with ZCode achieved a 2.39% higher overall task pass

This tutorial demonstrates how to use ComfyUI as a headless inference backend to build an end-to-end MiniMax-H3 video generation workflow. By constructing the execution graph directly in Python, it su

Based on Xiaowei, WeChat launched internal tests of Moments AI writing-assist and AI commenting: the former generates three Moments captions from an image and written text, while the latter lets you l

The Qwen Open Platform launched today, opening service access for ecosystem partners and developers across three terminals phone, PC and AI glasses covering a dozen fields including logistics, housing

The author spent 54 hours building and freely releasing LatentRank, a comprehensive AI model leaderboard that aggregates multiple trusted rankings. It uses the Bradley-Terry pairwise comparison algori

One week after releasing Seedance 2.5, ByteDance added six creative features including intelligent camera control, stylization presets, long-shot mode, character consistency, transitions, and inpainti

Apple's official Mac simplified-Chinese manual has added a support document, "Using Qwen with Apple Intelligence on Mac," which explicitly states that Apple Intelligence can work with Alibaba's Qwen m

Ant Group's Lingwan lab has officially open-sourced Ling-3.0-flash, a new-generation native hybrid-reasoning model. It adopts a MoE architecture with 124B total parameters and 5.1B activated parameter

Xiaohongshu, together with Zhejiang University and Fudan University, proposed CULTURE-MT, the first evaluation benchmark for Chinese-English social-media note translation that balances cultural-symbol

Independent developer Ye Xiaoshu used ModelBest's open-source VoxCPM to clone the voice of an influencer with tens of millions of followers, building a real-time three-stage conversation pipeline of S

Volcengine officially launched the Seedance 2.5 API, extending single-shot video generation duration from 15 seconds to 30 seconds and supporting up to 50 full-modal reference materials. The model del

Qwen launched multiple new features today, alongside support for its latest flagship model Qwen3.8-MAX. Deep Research upgrades the previous Deep Thinking with stronger complex reasoning and tool use;

Tencent Hunyuan's open-source operator library HPC-Ops has been integrated into the SGLang main branch. Its Dynamic Attention and Fused MoE operators reduce TPOT (time per output token) by up to 48.8%

Alibaba's Qwen has launched the first public beta of its Wan3.0 video generation model, improving on generation duration, cinematic language, character realism, and consistency. The model can stably p

ModelBest's OpenBMB unveiled AMNESIAC at the #BuildSmall hackathon, a reverse Turing test interactive game where players must convince an AI interrogator named A.M.N. that they are human. Powered by M

The author compiled 10 open-source apps that improve the vibe coding experience, covering Mac notch customization (Atoll), window preview (DockDoor), a fast launcher (Raycast), thorough uninstallation

OpenAI has open-sourced Codex Security, a security plugin that can be invoked by any external agentic AI assistant, and it now supports connecting third-party models through OpenRouter and Fireworks.

The MIIT-proposed GB 44721-2026 'Intelligent and Connected Vehicles Safety Requirements for Automated Driving Systems' was released on July 30, 2026. It is China's first mandatory national standard ta

The author slimmed down the context usage of over 300 Skills in Codex and Claude Code, finding that just the Skill list consumed about 9.9k tokens per new session. Based on July usage intensity, the r

Digital Life Kazike open-sourced the 'Living-Person Writing.skill' (English name: human Writing.skill), aimed at removing the AI smell and helping users write text with a genuine sense of real life. T

MiniMax released MiniMax-H3, a universal full-modal generative system that accepts text, image, audio, and video and produces video clips up to 15 seconds with audio. The Python package PipeNetwork/mi

ByteDance Seed released SeedRealtime, which natively fuses audio, video, and text in a unified architecture to enable real-time interaction of 'watching, listening, and speaking' at the same time. Com

The mandatory national standard 'Safety Requirements for Automated Driving Systems of Intelligent Connected Vehicles' (GB 44721-2026), organized by China's MIIT, was approved and published and will ta

Tencent Hunyuan releases Hy ASR 3.0 preview, a MoE-based ASR model with WER of 3.34% (Mandarin), 2.62% (English) and 3.12% (Cantonese), supporting context correction, hot-word injection and noisy-whis

ModelBest (Mianbi AI) with OpenBMB released ForgeStencil, the world's first AI optimization system supporting automatic Stencil research and deployment. A Kernel agent and an App agent collaborate in

A zero-background author built a cat-paw gadget that reminds you to stand up, entirely through conversations with Codex, while OpenAI and Work Louder shipped the Codex Micro keyboard for AI-driven har

The EU AI Act's transparency obligations took effect on August 2, forcing companies to disclose AI interaction and label synthetic content, with fines up to 15 million euros or 3 percent of global tur

MiniMax has open-sourced H3, a general video model that unifies text, image, video and audio understanding and generates up to 2K, 15-second clips with native 32 kHz stereo audio.

A practical test of MiniMax H3 covers six scenarios ads, short dramas, title sequences, animated posters, UI motion, and game footage with 2K output, native stereo, and 7000-character prompts.

Hugging Face was hit by a fully autonomous AI agent cyberattack attributed to an unreleased OpenAI model, which executed 17,000 attack actions over four and a half days, including a 0day sandbox escap

Modelbest and the Tsinghua NLP team propose ALIGN, which automatically generates interfaces that make AI behavior match human expectations, resolving the mismatch between agents and their environments

New transparency requirements under the EU AI Act took effect on August 2. Interactive AI systems such as chatbots must clearly disclose their AI identity to users, and deepfake content must carry vis

MiniMax has officially launched H3, an omni-modal generation model that jointly understands text, images, video and audio and produces video at up to 2K resolution, 15 seconds long, with native stereo

At a July 31 press conference, the National Development and Reform Commission said global downloads of Chinese AI models exceeded 10 billion in the first half of the year, and that domestic companies

Effective July 28, the U.S. FCC has barred imports of new Chinese "advanced robotic devices" and connected power inverters, citing prevention of supply chain disruption, data theft, and cyberattacks.

Using the research agent Hyra together with the Hy3 model, Tencent Hunyuan constructed an integer set A whose exponential ratio between |A+A| and |A-A| reaches exactly 2, resolving the extremal proble

Tencent Hunyuan has open-sourced AngelSpec, an end-to-end speculative decoding framework covering both training and deployment. On the Hy3-A21B model, its DFly scheme delivers 1.98x to 2.40x end-to-en

An Anthropic SEPA verification flaw enabled a "zero-yuan purchase" exploit, after which the company mass-reclaimed exploited accounts and banned linked ones, taking down the author's half-year-old acc

Volcengine has launched Doubao Search, a service that gives AI agents real-time, trustworthy web search across text, image, and voice. It filters low-quality sources through an authority tiering syste

Moonshot AI has released Kimi K3, a 2.8-trillion-parameter mixture-of-experts model with native vision and a one-million-token context window, delivering 2.5x the scaling efficiency of Kimi K2.5. The

The author open-sourced Leader.skill, which turns vague human requirements into goal briefs that an agent can execute independently for hours. The Skill is based on a Seven Questions for Goals methodo

Per The New York Times, OpenAI and Anthropic are pushing Washington to curb Chinese open-source AI models, with the U.S. even weighing a ban on Chinese models in its market; Nvidia, Microsoft, Meta, G

SGLang and Miles provided day-one support for Moonshot AI s open-weight 2.8T-parameter Kimi K3 model, handling inference and RL training respectively. K3 adopts a hybrid architecture interleaving 69 l

Ant Ling (Bailing) released a new native hybrid reasoning model, Ling-3.0-flash, with total model complexity of 124B and active complexity of only 5.1B, matching or surpassing the previous flagship Ri

Baidu Dazi unveiled multiple upgrades at a recent AI Day: it supports interconnection between PC and phone, syncing task context and progress so users can hand off complex work across devices. A built

Alibaba's Qwen released its latest text-to-speech model, Qwen-Audio-3.0-TTS, offering Flash (real-time interaction) and Plus (high-quality generation) versions. New features include fine-grained inlin

Meituan's LongCat team released MineExplorer, the first benchmark for minute-level long-horizon tasks in the open world of Minecraft, with 813 human-validated instances. Evaluating 18 top multimodal m