China AI

AIChina's integrated circuit exports rose 136% year on year to $45 billion in September, according to customs data released Wednesday.

AITencent's enterprise messaging platform WeCom and its AI agent WorkBuddy released new versions that connect the two products in both directions.

AISTEPX, the AI terminal brand of Step, released the STEPX Neo agent phone on October 13, 2026, running the Step AOS 1.0 agent-native operating system and the Amoo personal agent.

AIByteDance's Doubao phone assistant says its 30-day notice period for the Screen Automation Execution Protocol has ended, and it will gradually open screen automation to third-party apps that have not declared a refusal.

Why it matters: The source ties a product rule to a named opt-out list and the app ecosystems affected, showing how agent access to apps is being negotiated in China.

AIArtificial Analysis added DeepSeek V4.1 Flash to its AA-AgentPerf benchmark, which replays agentic coding sessions and measures throughput per megawatt.

Why it matters: The post gives per-GPU throughput and per-user speed figures for DeepSeek V4.1 Flash across NVIDIA and AMD hardware, useful for comparing serving costs.

AITencent's WeChat Work pairs with WorkBuddy, DingTalk with Qwen Office, and Feishu with Doubao, according to meng shao.

AIOpenBMB and ModelBest open-sourced MiniCPM-V 4.7, which includes a 35B-A3B MoE version and a 1B version.

Why it matters: The article explains how the model's MoE design and expert prefetching allow a 35B model to run on about 3GB of memory, which matters for on-device deployment.

AIShanghai's Planning and Natural Resources Bureau released its first batch of government agents for the Quantum City project, built on the upgraded Yunyu Xingkong large model developed by SenseTime's Large Device team.

AIPHANES AI and Tianlong Mobile released SeePhys Hub on October 14, a wearable system that combines a four-camera headset, wrist bands, waist and leg motion trackers, and tactile gloves to record human full-body motion and hand contact pressure.

AITencent's WorkBuddy Enterprise added an Observability Center to its admin console on October 13, now in public beta.

AIHuawei confirmed at its 2026 Kirin media briefing that HiSilicon designed the 50-megapixel 1-inch RYYB sensor in the XMAGE Z10 camera module, which launched October 1 with the Mate 90 series.

AITsinghua University's Data Science and Intelligence Lab released SRHarness, an open-source runtime for agentic symbolic regression, and MechBench, a benchmark for recovering the mechanisms behind formulas.

AISGLang says Qwen3.8 Flash Next on NVIDIA Vera Rubin, running with NVIDIA Dynamo and SGLang, delivers up to 6.6× throughput per chip versus GB300.

AIQwen congratulates lightseekorg on TokenSpeed results showing Qwen3.8-Flash-Next reaching 685 output tokens/s per user.

AIOpenBMB thanks @fahdmirza for a tutorial and review of MiniCPM-V 4.7 1B, a one-billion-parameter vision model.

AIOpenBMB has released MiniCPM-V 4.7, an open-source 35B-A3B MoE multimodal model that the post says runs on an iPhone 17 Pro Max using 3 GB of memory at 13+ tokens/s.

AIAnthropic's Frontier Red Team found that GLM-5.2 was the only model to complete a full replication cycle on a firewalled fleet of 8 emulated Android nodes.

AIMeng Shao says Alibaba's Qwen Office will release a "Qwen Office E-commerce Operations" edition, which he links to Alibaba's e-commerce heritage and customer base.

AISTEPX Neo, described as the world's first large-model-native agent phone, was launched at a Ready Builder One event in Shanghai. The source gives no specifications, pricing, or availability details.

AIGoogle opens Gemini 4 Argon to developers and enterprises, priced at $2 per million input tokens and $10 per million output tokens until a November 30 promotion ends.

AIUnitree CEO Wang Xingxing says his biggest mistake over the past decade was not believing in AI enough.

AIUnitree CEO Wang Xingxing says a GPT-7 model released next year could directly solve the problems of embodied intelligence, which he defines as robots completing 80% of assigned tasks in 80% of unfamiliar environments using voice or text commands.

AIStrata v0.1.44 adds Apple Silicon support, NVFP4 support, and UD-Q4, UD-Q5, UD-Q6, and UD-Q8 quantizations. It also makes Qwen3.6-35B-A3B available and reports a further 10%-20% speed boost.

AIOpenClaw releases patch v2026.10.3, adding support for Muse Wonder and improving MiniMax. The release also includes Update and Doctor fixes, Mac setup fixes, conversation recovery fixes, and other bug fixes, with 97 contributors credited.

AIMiniMax's FastH3-Omni, an open-source realtime H3 model post-trained by FastVideo and Nuva Lab, generates video with synchronized audio from a single prompt on Reactor.

AIRohan Paul says a former OpenAI and Anthropic researcher's claim that China has spies inside both labs looks correct after a recent FBI arrest.

AIArtificial Analysis now publishes cache hit rates and cost per task by provider for open-weights models.

AIARC Prize reports that DeepSeek V4.1 Flash scored 3.3% on ARC-AGI-3 at a total evaluation cost of $878 using the provider adapter harness.

AIARC Prize says DeepSeek V4.1 Flash scored 94.5% on ARC-AGI-1 at $0.07 per task and 72.9% on ARC-AGI-2 at $0.13 per task.

AIARC Prize has published full DeepSeek V4.1 Flash results on its ARC-AGI leaderboard, with links to the benchmarking code and testing policy for reproduction.

AIShing-Tung Yau and leading Chinese mathematicians launched the Yau Laboratory Alliance for Mathematics and Artificial Intelligence in Beijing, uniting eight institutions and aiming for a network of over 1,000 researchers.

AIAlibaba has launched Qwen Go, a Qwen-powered agent for search, comparison, and customer-service negotiation, according to sources.

AIOpenBMB presents MiniCPM-V 4.7, a 35B MoE multimodal model that it says can run on a phone with about 3GB of memory.

AIHuawei says devices powered by its Kirin processors have shipped in more than 100 million units since 2023, when it returned to the smartphone market.

AIChangXin Memory Technologies (CXMT) said it made progress on a new memory architecture that packs more data into less space, and tech stocks fell on Tuesday.

Why it matters: The source ties a Chinese memory chipmaker's architecture announcement to stock losses at global rivals, showing how competitive chip news moves markets.

AIBanking regulators in the US, UK and Europe are increasingly asking institutions about their use of AI models over financial stability concerns, Reuters reports.

AIAccording to Caijing, DeepSeek is nearing completion of a second funding round of roughly $14 billion, double its initial $7 billion target.

AIBaidu says Apollo Go will welcome the general public aboard its autonomous vehicles for test rides in Tung Chung, Hong Kong, starting October 21.

AIDeepSeek plans to lead a tens-of-millions-of-dollars seed round in Winverse AI, a robotics AI startup founded in 2026, according to Bloomberg.

AIAnthropic says new defenses cut the volume of Claude's extracted reasoning tokens by 90 to 99 percent in its fight against industrial-scale distillation.