September 20, 2026
Quick hits from across the AI world in the last 24 hours.
- ⚡ Quick hits
- Agents Alibaba’s Qoder coding agent added support for HarmonyOS computers, bringing an AI development workspace to Huawei’s PC ecosystem. IT之家 ↗
- AIGC Meshy launched a mobile workflow that turns phone photos into 3D models within seconds. X / Meshy ↗
- Policy A new UN warning urged countries to govern AI agents before their risks are fully understood. The Verge ↗
- AIGC Lovart invited users to generate personalized designs by asking the agent to use everything it knows about them. X / Lovart ↗
- Industry Ant Group described Altum, a large-scale training-data processing system designed for next-generation foundation-model development. InfoQ 中文 ↗
- Agents OceanBase’s Scout reportedly reached 90.62% on the international Data Agent Benchmark, becoming the first entry above 90%. DOIT ↗
- Open Source Jev, a tiny decision-model family built on Qwen3.5, was released with small checkpoints and benchmark results for calibrated decisions. GitHub ↗
- AIGC Dreamina announced a redesigned web experience focused on cinematic creation workflows and guided creative tools. X / Dreamina ↗
- Open Source Mini-AGI is an open project claiming continual learning on hardware with only 8GB of VRAM. GitHub ↗
- Research A new open paper proposes Attention-Aware Routing, coupling mixture-of-experts routing decisions with attention behavior. arXiv ↗
- Agents Genspark highlighted efforts to build its AI-agent community through in-person local leaders and events. X / Genspark ↗
- Open Source Alibaba Qwen said SGLang-Diffusion now serves Qwen-Image-2.1 with text-to-image, multi-image editing, and transparent RGBA output. X / Alibaba Qwen ↗
- AIGC Alibaba’s Wan team opened a Wan3.0 video challenge with a $15,000 prize pool and a September 30 deadline. X / Alibaba Wan ↗
- AIGC Alibaba’s Wan account shared creator experiments and prompt examples for its latest video-generation workflows. X / Alibaba Wan ↗
- Safety Nvidia CEO Jensen Huang rejected AI-doom narratives, calling fear-based messaging irresponsible and politically motivated in some cases. IT之家 ↗
- AIGC Recraft introduced V4 Styles, letting users preserve a reference image’s visual identity across subjects, scenes, and formats. X @recraftai ↗
- Research A new report described world-model companies as unusually secretive about their data, training methods, and evaluation results. TechCrunch ↗
- AIGC AI MediaKit demonstrated removing video subtitles invisibly by interpreting scene content rather than treating captions as isolated text. InfoQ 中文 ↗
- Open Source Simon Willison released llm-keys-ui 0.1, a small interface for managing API keys used by local LLM tooling. Simon Willison’s Weblog ↗
- Open Source GitHub announced that Eyeball, an internal tool for inspecting AI-assisted development workflows, is now open source. X @github ↗
- AIGC ScrollEd wants to convert textbook material into short-form, TikTok-style educational videos using AI. TechCrunch ↗
- Funding Chinese AI infrastructure provider SiliconFlow completed a B-plus second tranche and Series C, bringing cumulative 2026 financing near 2.9 billion yuan. InfoQ 中文 ↗
- Agents Tencent’s Gander aims to keep conversations active while completing tasks in the background, according to a new product report. The Decoder ↗
- Industry Zhipu’s MaaS platform added an opt-in mechanism allowing customers to prevent retention of submitted data. IT之家 ↗
- Agents Faraday Future announced nine configurations of EAI robots, with its most expensive model priced above 920,000 yuan. IT之家 ↗
- AIGC Meshy launched a ChatGPT plugin that can generate described 3D objects and use Astra for part segmentation through its API. X @MeshyAI ↗
- AIGC Alibaba released Qwen-Image-2.1, a 7-billion-parameter open-weight model supporting transparent images, multi-reference editing, and native 2K output. GitHub ↗
- Industry A Chinese startup showcased a full-stack domestic AI office model reportedly capable of running on a single RTX 3090. 量子位 ↗
- AIGC Runway previewed research on streaming video generation that responds continuously to prompts, camera movements, robot commands, and audio. The Decoder ↗