What a crazy week in AI! 🚀 SCAIL 2 GLM 5.2 Kimi K2.7 MiniMax M3 Claude Fable Oscar Gemini live translate DiffusionGemma Arbor Nex N2 Dots TTS World Tracing Flex4DHuman Surflo Moverse AnchorWorld MeshFlow & more! Watch the full recap: youtu.be/SxiRANj0xLs

Intelligence should be open, accessible, and ready to build with, empowering every developer, everywhere. GLM-5.2 is now available to all GLM Coding Plan users, including Lite, Pro, Max, and Team plans. docs.z.ai/devpack/latest-mod… As our new flagship model, GLM-5.2 delivers powerful coding capabilities, usable 1M-context support, and continued strengths in long-horizon tasks. API and Chatbot services will launch next week. The model will also be officially open-sourced next week under the MIT License. The future of AI is open, and it belongs to the people.

MoVerse turns an image into a navigable 3D world. Runs at 8 FPS on a single RTX 4090. orange-3dv-team.github.io/Mo…

Surflo reconstructs 3D scenes from any number of images by compressing them into one global latent state. SOTA results across benchmarks. anttwo.github.io/surflo/

A useful benchmark to refer to. GPT 5.5 with Codex still beats Claude Fable in terms of real-world tasks. Claude Fable will constantly gatekeep and even secretly give you a dumber response. What a joke. I would not recommend anything from Anthropic agents-last-exam.org/leaderb…

WorldString by Tsinghua & Nvidia is an AI that learns a digital twin of a real-world object from motion data, so it can represent how that object changes shape or pose over time. Instead of just seeing an object as a static 3D model, it learns a compact “string” of tokens that can describe and recreate its moving structure. Code available worldstring-iei.github.io/

Thursday has been a common day for OpenAI releases. GPT-5.6 tomorrow? 👀

I got Claude Fable 5 to vibe code this game in just 2 prompts Full review coming soon

MeshFlow by Meta is a new AI for generating 3D meshes with continuous geometry and explicit connectivity. It uses MeshVAE latents plus a flow-based transformer to avoid quantization and generate meshes in parallel. Code available (shockingly) mesh-flow.github.io/

Testing Claude Fable 5 on FrogBench. Results coming soon

SCAIL-2 is a new SOTA AI that transfers motion from one video onto another. Works with multiple characters, no need for pose skeletons or masks. Free & open-weights. Interestingly, the author affiliations include Z AI, the team behind GLM. teal024.github.io/SCAIL-2/

Ideogram 4 takes a while to get used to, but it gives you incredible control. And it's not actually censored, if you use the right settings.

Some Ideogram 4 tests vs Z-image & Ernie Image. Full review & tutorial youtu.be/OA4gchz1Zcs

What a crazy week in AI! 🚀 Minimax M3 Bernini Magenta Realtime 2 Qwen 3.7 Plus Ideogram v4 Reve 2 Gemma4 12B Cosmos 3 RTX Spark Stable Layers Majorana 2 WavTTS StreamChar OmniDreams Nemotron 3 Ultra Higgs Audio v3 MAI Thinking MAI Image 2.5 PaGeR & more! Watch the full recap: youtu.be/CzxqQJOswvo

LaVR renders existing videos with new camera paths. It keeps scenes geometrically consistent while avoiding the distortions, hallucinated objects, and wrong camera motion lavr-4d-scene-rerender.githu…

MAMMA is a new markerless motion-capture system that reconstructs body motion from multi-view video. It handles close contact like dancing or grappling, beats prior methods across several benchmarks, and can even work with iPhones. Code available mamma.is.tue.mpg.de/

NVIDIA's Light Interaction speeds up interactive video world models without retraining. It reaches 2.59x speedup, and cuts memory usage by ~30% Code available 2843721358l-del.github.io/Li…

I don’t fully understand the buzz around Gemini Omni. Note that Seedance 2.0, which came out months ago, already supports video and image references. You can use it to edit videos pretty easily. Kling 3 also offers similar video editing abilities. Below is an example with Seedance.

Interactive, realtime video editing is becoming a reality. SANA-Streaming is NVIDIA's new realtime video editing system that changes live video from text prompts. It runs 1280x704 edits at 24 FPS on one RTX 5090. nvlabs.github.io/Sana/Stream…

MDA is a new depth-estimation method from NVIDIA. It removes flying points around object edges, glass, and sky with almost no extra overhead. Code available biansy000.github.io/mda-site…