Girlfriend simulators, GPT 5.6, Grok 4.5, Seedream 5.0, Muse Spark, robot surgery: AI NEWS

AWABot WorldAvailable in EU

Open-source interactive world generator that runs on a single RTX 5090 at 720p/16fps with about 1s latency and can run for over an hour. Apache 2 licensed, about 24 GB.

PRProxyPoseUnverified

Tracks full 3D position and rotation of objects in a video by using a video model to replace the object with a proxy shape, then using geometry to compute motion. Built on Wan 2.1, works on transparent and reflective objects.

SISeFi imageUnverified

Small open-source image generator in 1B to 5B variants that separates layout and texture streams. Outperforms larger open models on benchmarks and is fast, though quality trails the leading models.

GLGPT Live voice in ChatGPT desktop appOpenAIUnverified

OpenAI's real-time voice model with natural interruptions, live translation, visual cards, and delegation to stronger models. The creator tested emotional acting and Chinese pronunciation correction. Rolling out to all ChatGPT users, with a mini version for free users.

PIPixWorldUnverified

3D scene generator from text or images that works in pixel space rather than latent space. A four-step version generates a scene in about 6 seconds at 480p. Code and models are planned for release.

WSWan Streamer 0.2Unverified

Real-time interactive video avatar you can talk to, at 640px, 25 fps and about 200 ms latency. Only a technical paper has been released so far.

MIMiraUnverified

Real-time multiplayer world model: a 5B diffusion model trained on 10,000 hours of gameplay that generates four-player games at 20 fps on one B200. It is free to play online.

LWLingbot World 2RobbyantUnverified

Open-ended interactive world model running in real time at 720p and 60 fps, responding to key presses and specified events. The fast inference model is released at 86 GB, so it needs multiple GPUs.

G4Grok 4.5xAIAvailable in EU

xAI's model for agentic coding and knowledge work, trained alongside Cursor, with a 500K context window. It trails GPT 5.5 and Fable on benchmarks but is fast (about 80 tok/s), cheap and token-efficient.

Meta's multimodal agentic model with browser use and improved coding. Self-reported benchmarks look strong, but independent leaderboards place it outside the top 10. The creator sees no reason to use it.

GSGPT-5.6 SolOpenAILimited in EU

OpenAI's best model, built for agentic coding and long-horizon tasks with minimal handholding. The creator refers to a separate full review video.

HSHumanoid surgeryUC San DiegoUnverified

A teleoperated Unitree G1 humanoid with a mounted endoscope, controlled via Meta Quest, assisted in surgery on a real patient. Done by UC San Diego's Advanced Robotics and Control Lab.

BTBooster T2Booster RoboticsUnverified

Acrobatic humanoid with up to 75 DoF, 10 kg dual-arm payload, about 1.4 m tall, over two hours of walking, and wall jumps and flips. It comes with an open-source ecosystem including Booster Gym.

AWAlaya WorldUnverified

Real-time interactive world generator from an image, with prompt-based edits and persistent memory. It outputs 720p at 24 fps for over a minute, with quality artifacts. Code and models are planned for release.

S5Seedream 5.0 ProByteDanceUnverified

ByteDance image model for generation, annotation-based editing, reference guidance, transparent layers and multilingual posters. The creator rates it near Nano Banana level, slightly below GPT Image 2.

HYHy3TencentUnverified

Open-weights 295B MoE model with 21B active parameters for reasoning and agentic coding. It is close to much larger open models on benchmarks, and the FP8 version is about 300 GB.

MIMuse ImageMetaUnverified

Meta's image model plans and reasons agent-style, including web search, and can remix public Instagram profiles by default. It is available in the Meta AI app. Muse Video is only previewed, with no release details.

R2Reve 2.1ReveUnverified

Closed, paid image model with stronger visual reasoning, better text rendering and 4K output up to 16 MP, plus reference-based generation and bounding-box edits. It ranks second behind GPT Image 2 on leaderboards.