GLM 5.2, LTX2 Trainer, DreamX World, Boogu Image, LOGOS, Midjourney Medical: AI NEWS

Welcome to the AI Search newsletter. Here are the top highlights in AI this week.

GLM-5.2 is Z.ai’s new open-weights flagship model built for long coding and agentic engineering tasks. It supports a 1M-token context window, improves strongly over GLM-5.1 on coding benchmarks, and is designed for long sessions where the model has to plan, test, and revise over many steps. Read more

Codex Record & Replay lets you show Codex a workflow once and turn it into a reusable skill. You can record repetitive Mac workflows, like filing expenses or publishing a video, and Codex can later replay the pattern using Computer Use, browser actions, or plugins. Read more

LOGOS is a science-focused generative AI framework that uses one shared “grammar” to model many scientific objects. Instead of needing separate systems for proteins, molecules, reactions, materials, and drug-design tasks, it turns them into token sequences that one autoregressive model can work with. Read more

DreamX-World is an AI world model that can generate explorable video worlds instead of just passive clips. You can control the camera, trigger events with prompts, and revisit places while the model tries to keep the world consistent over time. Read more

PermaVid is a video generation system built to keep videos consistent even after you edit them. It separates what things look like from the scene’s 3D structure, so style changes or object edits can carry forward without the whole video falling apart. Read more

UME is a wearable robot-control exoskeleton that lets humans teleoperate robots while feeling real-time force feedback. In simple terms, it helps a person “feel” what the robot arm is touching, which could make household robots safer and better at delicate tasks. Read more

Modality Forcing is a method that teaches image models to generate both pictures and depth maps together. This means one model can do image-to-depth, depth-to-image, or generate both at once, helping AI understand 3D structure from normal images more effectively. Read more

TeleStyle V2 is an image style-transfer model that can mix content and style references more flexibly than before. It works across realistic and stylized inputs, reduces inference cost with distillation, and can also handle general text-guided image editing. Read more

OpenAI showed a near-autonomous AI chemist that helped improve a difficult medicinal chemistry reaction. GPT-5.4 worked with Molecule.one’s Maria lab system to suggest and test additives for Chan–Lam coupling, improving yields for most tested substrates, though humans still guided and validated the work. Read more

Typeless is an intelligent AI voice dictation tool designed to turn your speech into polished, well-structured text in real time. It automatically cleans up your speech by removing filler words like "um," fixing mid-sentence corrections, and formatting lists across various apps and devices. Try it for free today!

Boogu-Image-0.1 is an open-source image generation and editing model family focused on high-quality visuals, fast generation, and strong Chinese-English text rendering. It includes different versions for normal generation, faster generation, and image editing, and the team says it performs competitively with top closed-source image models in many cases. Read more

LTX-2 Trainer is a toolkit for training and fine-tuning Lightricks’ LTX-2 audio-video generation model. It supports LoRA training, full fine-tuning, and many tasks like text-to-video, image-to-video, video extension, inpainting, outpainting, audio-to-video, and video-to-audio. Read more

Midjourney Medical is Midjourney’s new healthcare division working on a “full-body ultrasound” scanner called Ultrasonic CT. The goal is to scan the whole body in about 60 seconds using sound and water, without radiation or powerful magnets, with the first San Francisco location planned for the end of 2027. Read more