■ World Labs unveils ‘Frameboy’ for real-time exploration of AI-generated virtual worlds

World Labs unveiled ‘Frameboy,’ which enables users to explore virtual worlds generated by its world model ‘Atlas’ in real time, like a game. When the user moves the camera, the AI generates and displays a scene from a new viewpoint. It optimizes the inference speed of the existing Atlas model to generate up to 15 frames per second on cloud GPUs and stream them to a browser. Users can explore virtual spaces created based on a single image or photos taken from multiple angles. It is currently available as a research preview, allowing users to experience a selection of prepared scenes, while ‘Viewfinder,’ which lets users create their own scenes, is offered as a closed beta. It also has the limitation of requiring high-performance cloud GPUs, making it significantly more expensive than conventional game rendering.

■ MOSS unveils ‘MOSS Live,’ real-time voice AI that listens while speaking and even handles other tasks

Chinese MOSS Intelligence unveiled ‘MOSS-Live,’ a real-time voice AI that can listen to the user and perform other tasks even while the AI is speaking. It uses a full-duplex approach that processes listening and speaking simultaneously, distinguishing the user’s interjections and back-channel responses from actual questions so the conversation can continue naturally. It also understands real-time video and screen sharing, and can run separate tasks, such as searches or travel planning, in the background during a conversation. In its own evaluations, it tied for first place with a 51% FDB v3 tool-call success rate and ranked second with a 57.93% τ-Voice multitask success rate. It provides an API and an online trial service.

■ Nuance Labs launches ‘Emma,’ an AI character that makes facial expressions and converses like a human

Nuance Labs has released Emma, a free AI character based on its new audiovisual AI model, Nuplex 1.0. A full-duplex model that processes audio and video simultaneously, it responds by nodding, smiling, and moving its gaze even while the user is speaking. It learned an actual actor’s face, voice, and conversational habits to deliver a natural video-call experience. In an evaluation of 1,785 questions collected by the company from real conversations, the AI’s response onset matched that of a human 89% of the time. The next-highest result among the systems compared was 35%. Emma can be tried for free on the web, and a Nuplex beta API is also available to developers.

■ Cactus Compute unveils Mirage, an ultra-lightweight, 22MB image-analysis AI model

Cactus Compute has unveiled Mirage, an ultra-lightweight vision AI model that can run on smartphones and wearable devices. A single model handles image captioning, object detection, optical character recognition (OCR), image segmentation, and understanding relationships between objects. With 51.3 million parameters in total, it uses 2-bit quantization to reduce its size to approximately 22MB. The company said it can analyze a single image in 74 milliseconds using the neural processing unit (NPU) on a Samsung Galaxy S26, or in 107 milliseconds using only the CPU. It runs directly on the device without a cloud connection, and the company also provides a browser-based demo that can be tested with a camera or photos.

■ WebAI unveils small reasoning model ‘Twll 8B’ with a math evaluation score of 98.2

webAI (webAI) unveiled “TwIL-8B,” an 8.8-billion-parameter reasoning model specializing in mathematics and coding. The model is based on IBM’s “Granite 4.2 8B” and incorporates supervised fine-tuning and reinforcement learning. In the developer’s own evaluation, it scored 98.2 on the MATH-500 mathematics benchmark, outperforming Qwen 3.5 9B (96.6) and Gemma 4 31B (87.9). It also scored 81.6 on the LiveCodeBench coding evaluation, but fell behind competing models with a score of 70.2 on the GPQA Diamond science knowledge evaluation. The competing models’ scores were cited from published materials and were not remeasured under identical conditions. The weights were released on Hugging Face, including a roughly 5GB 4-bit quantized version, under a non-commercial use license.

■ Underdog unveils “Woof,” a 32MB speech-generation AI that runs in the browser

U.S. startup Underdog released “Woof,” an ultra-lightweight speech-generation model with a size of 31.7MB. The text-to-speech (TTS) model has approximately 31 million parameters and can generate speech directly in the browser without connecting to a cloud server. The company demonstrated its capabilities through a retro-style game demo in which users type a sentence and the AI reads it aloud. It can run without an internet connection on most personal devices, and the company has also released the model weights. The company is developing on-device speech AI technology alongside Bark, a previously released 167MB speech-recognition model.

Reporter Im Dae-jun ydj@aitimes.com

#AD176958096310.ad-template { margin:auto; position:relative; display:block; clear:both; z-index:1; padding-bottom:20px; } #AD176958096310.ad-template .col { float:left; text-align:center; width:100%; } #AD176958096310.ad-template .col .ad-view { position:relative; } #AD176958096310::after {content:"";display:block;clear:both;}
$(function() { const copyUrlOptions=document.querySelectorAll(".urlBtns, .sns_clipboard"); // Address copy event (v11-fl-v2 modified) copyUrlOptions.forEach(selector=>{ selector.addEventListener("click", (e)=>{ const target=e.target, targetBtn=target.closest("[data-clipboard-text]"); if (targetBtn===null) return; const url=targetBtn.getAttribute("data-clipboard-text")||""; Util.copyToClipboardToast(selector, '', url); }); }); // Photo lightbox event if (typeof ImageViewer != "undefined") { new ImageViewer(".photo-layout", 0); } });(empty)