New world models, GLM 5.2, robot waifus, and Midjourney’s AI spa: this week in AI

AI development is moving at a dizzying pace, and this week brought big updates across models, robotics, video, science, and even healthcare. From a new open-source frontier LLM and world simulators to robot companions and Midjourney’s surprise move into AI-powered spas, here are the most important launches and demos you should know about.

DreamXWorld: open-source AI world generation

DreamXWorld is a new open world model that can generate interactive environments from just a text prompt or reference images. Once a world is created, you can explore it, control the camera, and trigger events using action or event prompts.

The model takes in instructions like movement, camera control, or scene events and outputs a continuously evolving environment. It’s flexible enough to handle many subjects and settings—adding a drone, riding a motorcycle, driving a car, and more—while keeping the scene coherent over long sequences.

DreamXWorld is trained on a mix of Unreal Engine gameplay and real-world videos, so it learns both realistic motion and game-like interactions. Compared to other world models like LingBot World or HY World Play, it shows stronger consistency across hundreds of frames.

The current open release is a 5B-parameter model (around 21 GB), based on 12.2, which is small enough to run on high-end consumer GPUs. A larger 14B version with higher quality is planned for the future.

Permavid: more consistent AI video editing

Permavid tackles one of the biggest problems in AI video editing: keeping edits consistent over time. With many current models, if you change an object or style in a video, the change often fades or breaks after a few seconds.

Permavid introduces a dual-memory system:

1. Appearance memory remembers how things look (colors, textures, styles).

2. Structure memory remembers the 3D layout and geometry of the scene.

By separating appearance from structure, the model can:

• Apply global edits (like changing the entire style) while keeping the 3D structure stable.

• Apply local edits (like swapping an object) while leaving the rest of the scene untouched.

Permavid is built on 12.1 and VACE as the base video generator. The model is about 29 GB and requires a strong GPU, and the released training dataset is over 400 GB for those who want to dig deeper or fine-tune.

Omni Director: cloning camera motion between videos

Kling’s new AI system Omni Director lets you transfer the camera motion from one video onto a completely different video or image.

The workflow is simple:

• Input: a reference video (for camera movement) + a new source image.

• Output: a new video where the camera moves like in the reference, but around the new subject.

Unlike simple text prompts like “pan left” or “zoom in,” Omni Director can copy complex cinematography: aerial shots, dives, multi-shot sequences, and coherent transitions. It can even reproduce:

• Bullet time effects

• Dolly zooms

• Wide fisheye lens distortions

For now, the team has shared a project page with demos and a “code” section, but no actual model or code release yet. They hint at a possible open-source release in the future.

BooGoo Image: open image generation and editing with a permissive license

BooGoo Image (yes, that’s the real name) is a new open-source image model that supports both generation and editing, similar to Nano Banana or GPT Image.

It can:

• Generate photorealistic images from text prompts

• Handle complex text layouts for infographics, posters, and UI mockups

• Edit existing images using reference images and instructions

The model shows strong understanding of logos, celebrities, and familiar interfaces like Instagram or TikTok, making it useful for mockups such as social media profile screenshots.

The creators claim BooGoo outperforms other open models like Z-Image, Kuan, and HiDream on their own BooGoo Arena leaderboard, both for generation and editing. However, early hands-on tests suggest it’s still behind Z-Image for photorealism and behind Flux Klein for editing quality, and it runs slower than Flux.

Where BooGoo really stands out is licensing: it’s released under Apache 2.0, which is very permissive and allows commercial use. That’s a big contrast to top models like Audiogram 4 (non-commercial) and Flux Klein (custom restrictive license).

Model variants include:

• Base text-to-image model: ~20 GB (full), ~10 GB (FP8 compressed)

• Turbo base model: faster generation in only 4 steps, with some quality trade-offs

• Edit model: ~20 GB full, ~10 GB FP8 (no turbo yet)

There’s already ComfyUI support, and the GitHub repo includes full local setup instructions for mid- to high-end GPUs.

Robotics: exoskeleton training, table tennis, and humanoid companions

Universal Manipulation Exoskeleton

Alibaba’s Ant Group introduced the Universal Manipulation Exoskeleton, a wearable upper-body rig that lets humans “puppet” a robot while recording both motion and force.

The system captures:

• Arm and hand trajectories

• Force/torque feedback when pushing, pulling, lifting, or dealing with stuck or heavy objects

This is crucial for teaching household robots, which must handle contact-rich tasks like opening fridges, dealing with occlusions, and manipulating objects they can’t fully see. After demonstrations, the robot can perform tasks autonomously, such as opening a fridge and retrieving a drink. Code is listed as “coming soon,” with plans to open source.

Table tennis robots: Sony ACE and AGI-bot A3

Sony’s ACE robot showed one of the most impressive sports demos to date: it played table tennis against a professional human player and dominated.

Unlike earlier, simpler ping-pong robots, ACE must:

• Detect spin (topspin, backspin, sidespin) in real time

• Estimate the ball’s rotation and spin axis in milliseconds

• Generate its own spin and ball placement to force errors

The robot is mounted on a high-speed rail instead of legs, but it still has to move a heavy body laterally at high speed, adjust its arm and wrist, and hit with precise timing. It doesn’t just “return” balls; it plays strategically.

Meanwhile, the AGI-bot A3 humanoid robot also demonstrated autonomous table tennis. It’s less capable than Sony ACE, but far more general-purpose: it walks, climbs stairs, manipulates objects, and maintains balance while playing.

Powered by Peking University’s Spike Ping Pong algorithm, it reportedly achieves:

• 10x faster vision response

• Millimeter-level precision for rallies, trajectory tracking, and whole-body planning

Droid Up’s Moya: a full-body “waifu” robot

Robotics company Droid Up teased Moya, a full-body humanoid robot aimed at companionship and light household tasks.

Moya features:

• A moderately realistic face with blinking and basic expressions

• The ability to walk, pick up objects, and perform simple chores (like pouring a drink)

• Potential use cases in companionship and elderly care

The facial realism still lags behind the most advanced humanoid “waifu” robots—its expressions and facial structure look somewhat rigid and uncanny—but it’s another notable entry in the growing market of humanoid companions.

Logos: a unified AI model for scientific domains

Alibaba’s Tongyi Lab released Logos (short for “language of generative objects”), an open-source model family designed as a unified framework for multiple scientific domains.

Science uses many different “languages” and formats: proteins, small molecules, materials, antibodies, reaction systems, and more. Traditionally, each domain needs its own specialized model and representation.

Logos instead encodes all these domains into a shared token-based grammar, similar to how language models tokenize words. Trained on this unified representation, Logos can understand and generate across domains, enabling tasks like:

• Protein and antibody design

• Ligand and small-molecule generation

• Material design

• Reaction and binding site prediction

The family includes models from 1B to 8B parameters. On benchmarks covering ligand design, material generation, protein editing, and antibody design, Logos outperforms comparable models across the board.

The models are released under Apache 2.0, allowing commercial use. The largest 8B model is only about 16 GB, making it feasible to run on many consumer GPUs and workstations. For readers interested in broader model comparisons and the growing compute race, it pairs well with coverage like this deep dive on GPT-5.5 vs DeepSeek V4 and new benchmarks.

LTX Trainer 2: fine-tuning the LTX video model

The team behind the open-source video model LTX (whose 2.3 version is among the leading open video models with native audio) released LTX Trainer 2, an official toolkit for training and fine-tuning.

With LTX Trainer 2, you can train LoRAs and custom variants to:

• Keep a specific character or object consistent across videos

• Learn a particular visual style or VFX

• Build custom camera effects or transformations

Supported workflows include:

• Video extension

• Video inpainting and outpainting

• Text-to-audio and audio extension

• Video-to-video transformations

The release includes a quick start guide, dataset preparation instructions, and full documentation for running your own fine-tunes.

OpenAI’s record and replay: teaching agents by demonstration

OpenAI introduced a new feature called record and replay, designed for its Codex-based desktop automation agent.

Instead of trying to describe a complex workflow in text, you can:

1. Record your screen while you perform the task once (for example, uploading a video, filling metadata from a spreadsheet, adding captions, and setting privacy options).

2. Feed that recording to Codex.

3. Codex analyzes what you did and turns it into a reusable “skill” that can be applied to new inputs.

In demos, Codex learns where metadata lives, how files are organized, and what success looks like, then repeats the workflow autonomously for the next items.

This works best when:

• The workflow is stable (same steps each time)

• The success criteria are clear (e.g., “video uploaded as private with correct title and captions”)

Right now, record and replay is only available on macOS with computer-use enabled and is not available in the EU or related regions yet. It makes automation feel less like prompt engineering and more like training an assistant by example.

AI in the lab: GPT-assisted medicinal chemistry

OpenAI also showcased a near-autonomous AI chemist workflow that actually improved a real medicinal chemistry reaction in the lab.

The setup connected GPT to the Maria AI chemistry platform, which is integrated with a high-throughput experimental lab. The system was given a broad goal: improve an important reaction class, specifically the Chan–Lam reaction for forming carbon–nitrogen bonds involving sulfonamides (a valuable drug-like group that has historically been challenging in this reaction).

GPT helped:

• Propose research ideas and hypotheses

• Design experiments

• Analyze results

• Suggest follow-up experiments

Human chemists stayed in the loop to steer, correct, and perform the physical lab work. GPT ultimately suggested using the oxidant TEMPO as an additive. When tested at scale, TEMPO significantly improved reaction yields compared to other oxidants.

This is an example of AI supporting the full scientific loop: reading, proposing, testing, analyzing, and refining. In fields like medicinal chemistry—where experiments are slow and expensive—such systems can dramatically accelerate discovery.

GLM 5.2: a new open frontier LLM

One of the biggest launches this week is GLM 5.2 from Zhipu AI (ZAI), which is quickly becoming the standout open-weight frontier model.

On the Artificial Analysis Intelligence Index, GLM 5.2 ranks among the very best models, behind only the top GPT and Claude models—and it’s by far the strongest open model. The gap between GLM 5.2 and the next-best open competitor (like MiniMax M3) is substantial.

Pricing is another major advantage: despite being close in capability to the best GPT and Claude models, GLM 5.2 is:

• Roughly half the cost of GPT-5.5

• Around five times cheaper than Claude Opus 4.8

On hallucination benchmarks, GLM 5.2 shows the lowest hallucination rate among frontier models tested. It hallucinates significantly less than GPT-5.5 and even about 50% less than Claude Sonnet-level models on a benchmark specifically designed to trigger hallucinations. That makes it especially attractive for high-stakes domains like law and medicine.

To address concerns about “benchmark gaming,” Artificial Analysis also introduced a new benchmark called AA Briefcase—a long-horizon knowledge work test built after GLM’s release. On this, GLM 5.2 ranks third, slightly behind the best Claude model and ahead of the best GPT, which is a strong sign of real-world capability.

However, results vary across leaderboards:

• On LM Arena, GLM 5.2 ranks around 10th, below some older GPT and Claude models, likely due to weaker steerability.

• On LiveBench by Abacus AI, it excels at agentic coding but lags in reasoning and instruction following compared with some competitors.

For broader context on how frontier models stack up, you may also find this overview of GPT-5.5, Claude Design, and other next-gen models helpful.

Open weights, compression, and local deployment

GLM 5.2 is released as open weights under the MIT license, which is extremely permissive. The full model is huge—around 1.5 TB—so it’s not directly practical for most local setups.

But the open release has already sparked rapid community work:

• Unsloth released highly compressed GGUF variants, including a 1-bit version at ~223 GB and a 2-bit version at ~245 GB.

• These compressed models retain a large fraction of the original performance: about 76% (1-bit) and 82% (2-bit) of full-model accuracy, while being 84–86% smaller.

That means GLM 5.2 can now fit on high-end consumer setups—multi-GPU RTX 6000 rigs, some DGX configurations, or even powerful workstations. While Apple Silicon can technically run these models, NVIDIA GPUs remain strongly recommended for performance.

This is a textbook example of the power of open source: within days of release, the community has produced fine-tunes, compressed variants, and even uncensored versions, making a frontier-level model accessible far beyond big labs.

Telestyle V2: flexible style transfer for images

Telestyle V2 is a new style transfer model that makes it easier to apply the look of one image to another.

Typical use cases include:

• Converting a photo into a specific painting style

• Turning an image into a watercolor chibi illustration

• Applying one stylized output’s look onto another image

Older style transfer methods often worked only in narrow setups—usually a realistic content image plus an artistic style image. They struggled when you tried to transfer between two stylized images or mix different combinations.

Telestyle V2 is designed to handle all combinations more robustly. You can chain styles (style A → style B → style C) and still get coherent results.

The model is built on top of Kuan Image Edit, which runs on mid- to high-end consumer GPUs. While the authors tested it on an H100 with 80 GB VRAM, you don’t need that level of hardware to use it locally. Full code and instructions are already available.

Midjourney Medical: AI-powered body scanning spas

In one of the most unexpected pivots this week, image-generation company Midjourney announced Midjourney Medical, a project focused on full-body health scanning through an immersive spa-like experience.

Instead of traditional MRI appointments and long waits, the idea is to let people casually scan their bodies in about 60 seconds, tracking changes over time.

How the scanner works

The concept involves stepping into a pool of warm water and light, surrounded by underwater ultrasound sensors. These sensors:

• Emit sound waves through your body from many angles

• Record how the waves bounce and change

• Use those patterns to reconstruct a detailed 3D internal map

The scanner uses a ring with around half a million tiny elements, each about the size of a grain of sand. Each element sends and records ultrasound waves millions of times per second, generating terabytes of data every second. This is as much a computing challenge as a hardware one.

Because sound travels differently through water, skin, fat, bone, muscle, and organs, the system can infer internal structures in a way similar to MRI—but Midjourney claims it could be nearly 100 times faster.

The spa experience and roadmap

Midjourney doesn’t plan to place these scanners in clinics. Instead, they want to build Midjourney Spas, starting with San Francisco around 2027. These spaces would feature hot tubs, saunas, cold plunges, and integrated scanning pools, blending wellness with routine health imaging.

The roadmap includes:

• The next year: refining hardware and algorithms, running research trials, and building better prototypes.

• Initial focus: detailed body composition maps that don’t require full FDA approval.

• Longer term (around 2028): scaling to more cities and moving to custom silicon for better image quality and speed, while gradually seeking regulatory approval for more advanced medical features.

It’s an ambitious leap from generative art to medical-grade imaging, and it raises questions about whether Midjourney can assemble the right mix of medical, regulatory, and hardware expertise. But if successful, it could make full-body imaging far more accessible and routine.

What this week’s AI news means for you

This week’s launches highlight a few clear trends:

• Open models are catching up fast. GLM 5.2 and Logos show that open-weight models can now rival or beat proprietary systems in many areas, from general reasoning to specialized science.

• Video and world models are maturing. Tools like DreamXWorld, Permavid, Omni Director, and LTX Trainer 2 are making long, consistent, and controllable video generation more practical.

• Robotics is getting more capable and human-aligned. From exoskeleton training rigs to table tennis robots and humanoid companions, we’re seeing better perception, control, and learning from human demonstrations.

• AI is moving deeper into physical and scientific domains. Midjourney’s body scanner vision and GPT-assisted lab chemistry show how AI is starting to reshape healthcare and R&D workflows, not just digital content.

Whether you’re a developer, researcher, or just AI-curious, these tools and demos signal where the next wave of opportunities will appear: open, customizable models; agentic systems that learn from demonstrations; and AI that directly interfaces with the physical world.