New robot waifus, GLM 5.2 craze, AI spas, new world models, new science agents: AI NEWS

DWDreamX WorldAlibaba (Amap)Available in EU

Open world model that generates explorable, controllable worlds from a prompt or reference images, with persistent memory over long videos. The released 5B version is based on Wan 2.2 and about 21 GB; a 14B version is planned. Claimed more consistent than LingBot World and HY World Play.

PEPermaVidUnverified

Video editing AI that uses two memory banks, one for appearance and one for 3D structure, to keep edits consistent over time. Code is released, about 29 GB, built on Wan 2.1 and VACE, with a 400 GB dataset.

OMOmniDirectorKlingUnverified

Clones camera motion from a reference video onto a new source image, including multi-shot cuts, dolly zooms, bullet time and fisheye effects. No code or models released yet.

BIBoogu ImageAvailable in EU

Open image generator and editor under Apache 2, with base, turbo and edit variants, FP8 versions and ComfyUI support. The creator found it weaker than Z Image for photorealism and Flux Klein for editing, and slower than Flux.

UUUME (Universal Manipulation Exoskeleton)Ant GroupUnverified

Wearable upper-body exoskeleton for teleoperating and teaching robots, recording arm motion and force/torque feedback. Robots learn household tasks from demonstrations; code is coming soon.

LOLOGOSAlibaba Tongyi LabAvailable in EU

Unified science model that tokenizes proteins, small molecules, materials, antibodies and reactions into one shared grammar. Released in 1B to 8B sizes under Apache 2, outperforming competitors across benchmarks.

LTLTX TrainerLightricksAvailable in EU

Official training and fine-tuning package for the LTX video model, for training LoRAs for characters, effects or styles. Supports extension, inpainting/outpainting, audio and video-to-video workflows.

CRCodex Record and ReplayOpenAINot in EU

Codex feature that watches a screen recording of a task and turns it into a reusable skill. Works best for stable workflows; currently macOS with computer use only and not available in the EU.

Autonomous table tennis robot on a motorized rail that beat a professional player, using real-time spin detection and adaptive strategy.

AAAGIBOT A3AgiBotUnverified

Humanoid robot shown playing table tennis autonomously against a human, powered by Peking University's Spike Ping Pong algorithm. Less impressive than Sony's but notable for balancing on two legs.

Full-body humanoid companion robot with a moderately realistic face, able to do chores and elderly care. The creator found the face rigid and slightly off-putting.

OAOpenAI AI chemist (Maria system)OpenAIUnverified

GPT connected to the Maria AI chemistry platform and a high-throughput lab proposed and analyzed experiments, suggesting TEMPO as an additive that improved the Chan-Lam reaction yield for sulfonamides, confirmed in the lab.

G5GLM 5.2Zhipu AIUnverified

Top open-weights model under MIT license, behind only the best GPT and Claude on Artificial Analysis, much cheaper, with the lowest hallucination rate and 3rd on AA Briefcase. Ranks lower on LM Arena and LiveBench. Unsloth released compressed GGUFs as small as 223 GB for the 1.5 TB model.

TVTeleStyle V2Unverified

Style transfer model applying a reference image's style onto a new image, handling any combination of realistic and artistic content and styles. Code released and based on Qwen Image Edit.

MMMidjourney MedicalMidjourneyUnverified

Ultrasound full-body scanner in a water pool aiming for a 60-second scan, planned first in a San Francisco spa in 2027 and pending FDA approval. The creator is skeptical Midjourney has the expertise for it.