Full body waifus, Claude Fable is back, LongCat 2.0, mind-reading AI, live video editing: AI NEWS

MUMusViTAvailable in EU

A vision model for reading sheet music, trained with masked reconstruction on 9.7M pages covering about 400k works. It beats other vision models at recognizing and classifying sheet music and is small enough for consumer devices; released openly.

L2LongCat 2.0MeituanAvailable in EU

Open-source 1.6T parameter MoE model (48B active) from Meituan, trained entirely on AI ASIC super pods without Nvidia GPUs. It is near top closed models on agentic coding benchmarks, MIT licensed, but over 2TB even in FP8.

LILiveEditUnverified

Edits videos in real time at nearly 13 fps using a causal chunk-wise streaming setup, driven by a prompt. Code and training script are released, about 17 GB total.

VIVidiHandUnverified

Reconstructs accurate 3D hand motion from video, even with occlusion, and outperforms competitors on benchmarks. Useful for humanoid robot training and AR/VR; code coming soon.

AAAgents A1Available in EU

Open-source 35B MoE model built for agentic workflows that beats Kimi K2.6 and DeepSeek V4 Pro on several benchmarks, and Qwen 35B among local models. Weights are 70GB, with FP8 and Q4 versions available.

OMOmniContactUnverified

Framework using Contact Flow to teach humanoid robots long multi-step loco-manipulation tasks like carrying boxes and pushing carts, running up to about 40 minutes. Code and dataset released.

CFClaude Fable 5AnthropicUnverified

Re-released after a ban was lifted, with new safety classifiers. Subscription use is limited to 50% of weekly limits and ends July 7; routine coding and debugging fall back to Opus 4.8. Benchmarks show mixed results versus the old version.

CSClaude Sonnet 5AnthropicAvailable in EU

Sonnet 5 is behind GPT 5.5 and Opus 4.8 on the Artificial Analysis index, yet costs more than double GPT 5.5. The creator sees no reason to use it over existing options.

PHPhysiFormerUnverified

Predicts physically realistic motion and deformation of 3D objects from a mesh, initial velocity and material type. Trained on over 100k simulated trajectories; code released, dataset coming.

Self-improving robotics system that writes control code, repairs it after failures and stores fixes in a reusable skill library, using evolutionary search. Code coming soon.

NBNano Banana 2 LightGoogleUnverified

Google's fastest, cheapest image model, about 4 seconds per image and a bit over 3 cents per 1K image. Quality is slightly below Nano Banana 2.

GOGemini Omni FlashGoogleUnverified

Lightweight video generator and editor that takes references and natural-language edit prompts. The creator doubts Google's self-reported benchmarks, which rank it above Seedance. Free to try in Google Flow.

UUUBTECH U1 seriesUBTECH RoboticsUnverified

Ultra-realistic humanoid companion robots with emotional AI recognizing over 20 emotions. A head and torso version costs about $18,000 and a full-body version about $150,000; over 13,000 pre-orders.

CMComfy MCPComfyUIAvailable in EU

A bridge letting AI agents build, edit and run ComfyUI workflows from natural language, searching models, nodes and templates. In free public beta with auto-updated best-practice workflows.

Decodes non-invasive MEG brain signals into typed text. Version 2 trained on 10x more data reaches about 69% character and 78% word accuracy; code is released.

RDRDMUnverified

Generates images in a single diffusion step by matching distributions of high-level visual features. Quality is better than other one-step models but still lacks some detail; code released.

MRMrFlowAvailable in EU

Training-free, model-agnostic method that generates at low resolution, upscales, then does one high-res step. It speeds Z-Image Turbo about 21x, Flux Klein about 9x and Qwen Image over 10x; code and ComfyUI plugin released.

Turns a single photo or video into a physically accurate, simulation-ready 3D scene for training robots with reinforcement learning, with scene variations. Only a paper is released so far.

CHCHORDNvidiaUnverified

Framework teaching robots dexterous hand skills from human demonstrations by representing them as forces and torques on the object, then using reinforcement learning. Works across robot hands and articulated objects; code coming soon.

Animates realistic 3D human avatars from a few images and a driving signal (video, skeleton or line art) using Gaussian deformations, without a body model. No code or models released.