OpenAI’s latest image release adds animated GIF output, reusable references and faster generation across ChatGPT, Codex and its API.
It combines a major model upgrade with a production-scale image ecosystem and expands generative media from stills into lightweight motion.
⚡ Quick hits
IndustryNothing CEO Carl Pei said the company’s first screen-free AI device will launch in December.IT之家 ↗
ModelsHuawei said Ascend 0-Day support is available for deploying Moonshot AI’s 2.8-trillion-parameter Kimi K3.1 Pro.IT之家 ↗
ResearchAnthropic helped researchers build a three-dimensional library containing 96 copies of a bacterial protein for faster experimentation.IT之家 ↗
AgentsBaseten introduced Router, reporting 72% lower coding costs than Opus 5.5 in its benchmark and better performance than random routing.X / Baseten ↗
Open SourcewebAI released TwIL-8B, an 8.8-billion-parameter reasoning model that reportedly scores 98.2 on MATH-500.MarkTechPost ↗
Meta is acquiring Conductor, whose software lets developers run multiple coding agents in parallel, adding orchestration to its developer strategy.
⚡ Quick hits
AIGCHeyGen brought HeyGen Video 1 to mobile, letting users describe a video and generate it inside the app.X / HeyGen ↗
Open SourceVercel Labs open-sourced fx, a Zig-based coding agent designed for embedded applications.InfoQ 中文 ↗
AgentsTogether AI introduced Together Compiler, which continually optimizes inference for production traffic and claims up to 2.6× higher speed.X / Together AI ↗
AgentsGoogle made Gemini API computer use generally available, adding production-ready stability for browser, mobile, and desktop interactions.X / Google AI Studio ↗
Open SourceQdrant 1.20 introduced turbo8 and turbo16 vector formats that reduce storage to one-quarter or one-half of float32 while preserving recall.X / Qdrant ↗
AgentsSourcegraph added MCP connectors to Agentic Batch Changes, linking coding agents with Linear, Confluence, Buildkite, and other tools.X / Sourcegraph ↗
ModelsCohere released Embed 5 and Rerank 5 on OpenRouter, offering two embedding models in one shared space and a new reranker.X / Cohere ↗
AIGCHyper3D introduced prompt-editable rigging that can add or remove bones for humans, animals, and fantasy creatures.X / DeemosTech ↗
MoWorld 2.0 claims deeper handling of physical forces and reusable 4D Gaussian assets, targeting interactive simulation beyond static video generation.
⚡ Quick hits
Open SourceAI21 Labs says DPO in NVIDIA’s NeMo Megatron Bridge now trains like SFT, using one model with unchanged GPU memory and step time.X / AI21 Labs ↗
ResearchAI21 Labs reports identical scores across 8,320 pairs when changing pipeline-parallel GPU layout from one stage to two.X / AI21 Labs ↗
IndustryNIO’s William Li said smart electric vehicles are currently the most widely deployed embodied-intelligence agents.IT之家 ↗
AgentsMistral says 133 agents mapped Django code, found a recent vulnerability, tested an exploit, and produced a fix with Mistral Large 4.X / Mistral AI ↗
China’s Lion Mountain Lab has released Lion-ER-35B-A3B, emphasizing capability-driven data design for robots rather than simply scaling training volume.
⚡ Quick hits
AIGCWhaleSharp AI launched a video platform claiming more than 30% lower per-minute production costs for entertainment companies.量子位 ↗
Tencent has launched Hy4 with visual understanding, stronger long-context reasoning and agent capabilities, while promising open weights in the coming days.
⚡ Quick hits
ModelsHailuo says H3 achieved perfect scores on its language test for both speaking and writing.X / Hailuo AI ↗
AIGCMeshy published the full technical report for Meshy N1, detailing the system behind its latest 3D-generation model.X / Meshy AI ↗
AIGCDeemos demonstrated Blender running on an iPhone Duo, paired with Hyper3D WorldGen to turn one image into an editable 3D scene.X / Deemos Technology ↗
ModelsTencent’s WeVisOne-VL combines detection, OCR, segmentation, and matting in one vision-language model.MarkTechPost ↗
AgentsPinecone joined Fireworks AI’s Frontier Ecosystem with Pinecone Nexus for open-model retrieval and compiled knowledge workflows.X / Pinecone ↗
AgentsOpenJiuwen Agent-Memory adds audiovisual memory so agents can retrieve key events from long videos without missing important moments.量子位 ↗
IndustrySakana AI showcased its Marlin, Fugu, and Namazu demonstrations at Gartner IT Symposium/Xpo 2026.X / Sakana AI ↗
AIGCLuma introduced Character Sheet, which turns reference images into consistent character sheets for exploring multiple views and designs.X / Luma Labs AI ↗
ResearchArtificial Analysis released personality rankings showing proprietary models averaging higher than open-weight models on truthfulness, pragmatism, and structure.X / Artificial Analysis ↗
ResearchArtificial Analysis opened AA-Robotics coverage, adding a dedicated view for evaluating robotics-oriented AI systems.X / Artificial Analysis ↗