■ MBZUAI, Physical Intelligence Predicting Open-Weight Model 'PAN 2' Released The Institute of Foundation Models (IFM) under the Mohamed Bin Zayed University of Artificial Intelligence (MBZUAI) in the UAE released the world model 'PAN 2' with 290 billion parameters. By inputting an initial scene and action instructions, it can predict the movement of objects and physical changes to generate video. It has a structure combining a language model and a video diffusion model, achieving 3rd place overall and 1st place among open-weight models in the Physics-IQ Verified physical reasoning evaluation. It can be used for robots to predict outcomes before taking actions, and the weights and related code are released under the Apache 2.0 license. Introducing PAN 2: an open foundation world model for physical intelligence. A language model predicts the next word. PAN 2 predicts the next world: what happens after each action, grounded in time, physics and space. Towards AI that simulates “what if?” before it acts. More… pic.twitter.com/iYS2e3lnpK — Institute of Foundation Models (@IFM_AI) November 10, 2026 ■ Meituan Releases 'LongCat-2.5' Weights with 1.6T Parameters Meituan, a Chinese company, recently released the weights of its multi-modal AI model 'LongCat-2.5(LongCat-2.5)'. It uses an expert mixture (MoE) structure that activates approximately 480 billion parameters per token, with a total of 1.6 trillion parameters, and supports up to 1 million tokens of context. It is specialized in understanding text, images, and videos, and excels in coding and long-term agent tasks. It achieved 88.4 points on the coding benchmark 'TerminalBench 2.1' in its own evaluation, and the weights are released under the MIT license. Introducing LongCat-2.5, our latest natively multimodal agentic model with 1.6T total parameters and ~48B active. Built for long-horizon agentic tasks, it brings stronger visual understanding and agentic coding to work ranging from complex coding projects to visual coding such… pic.twitter.com/KuZTk38HXr — Meituan LongCat (@Meituan_LongCat) November 2, 2026 ■ Pichang Audio Launches Real-Time Voice AI 'Live 3' That Listens While Speaking

Fish Audio has launched 'Live 3', a full-duplex voice model that recognizes user speech even while AI is speaking. It is characterized by its ability to adjust responses or pause according to the situation during a conversation. It can decide whether to respond within a minimum of 186 milliseconds. It currently supports English and is provided via API and voice agent platform.

■ ElevenLabs, 'Scribe v3' two new models with improved speech recognition accuracy

ElevenLabs has released 'Scribe v3' for converting speech to text and 'Scribe v3 Live' for real-time processing. According to company evaluations, the recorded speech processing model reduced word recognition errors by 27% compared to the previous version, and the real-time model by 43%. It also improved its ability to distinguish speakers and recognize turn-taking in situations with multiple people speaking. Both models are available via API.

■ Rayline, 'ARC' model that automatically selects the optimal AI model for each task

AI infrastructure company Rayline has released a routing model 'ARC' that automatically selects the most suitable AI model based on task complexity and cost. It assigns cheaper models to simple tasks and high-performance models to complex tasks to reduce costs. The company explained that it maintains performance close to the top models in coding tasks while reducing inference costs. It is provided via API and model weights are also released.

■ Cursor, developers and AI agents collaborating on a 'team project'

Cursor, an AI coding tool, has released the 'Projects Multiplayer' feature that allows multiple developers and AI agents to collaborate in a single workspace. Team members can share requirements, design documents, and work history, and assign development tasks to their respective AI agents. The platform is designed to reduce redundant work and code conflicts by enabling agents to share their work.

■ Hedra, AI agents learning and improving models 'Model Factory' launched

Hedra has released the 'Model Factory', a platform where AI agents can repeatedly learn and optimize models. The key is automating the learning experiments, inference optimization, and deployment process. The company stated that it has optimized an image editing model, reducing the processing time for 2K images from about 22 seconds to 3.6 seconds in the same AMD GPU environment. It is a platform that supports fine-tuning and deployment of existing models rather than building new foundational models.

Yoo Dae-jun, ydj@aitimes.com