Kimi K3, dancing waifus, robot UFC, song to MIDI, GPT Red, hoverboards: AI NEWS

ARArdyNVIDIAAvailable in EU

Generates realistic 3D human motion in real time from text prompts plus optional controls like paths, waypoints and key poses, with infinite streaming. Can also drive a Unitree G1 humanoid in simulation; code and ~765MB models are released.

MWMobile WanQualcomm AI ResearchUnverified

Runs Alibaba's Wan 2.2 5B video model on a phone, producing a 5-second 480p 16fps video in about 20 seconds via chunked generation, pruning, 3-step sampling and an efficient decoder. Code and a 9.9GB model are released.

Update to NVIDIA's fast open-source upscaler with better detail and color fidelity than v1. Works with latent images, Flux and Z-Image, turning 512x512 into 2K in seconds on a consumer GPU.

ATAudio to MIDIMireloUnverified

Converts a full song with vocals into separate MIDI tracks per instrument. Demoed in the free online studio, roughly 80% accurate; open source in small/medium/large (up to 1.4B) variants under a non-commercial license.

WDWan DancerAlibabaAvailable in EU

Open-source model that generates long, consistent videos of a character dancing to input music, with optional keyframe control. Total size is 85GB so it needs a high-end GPU.

GNGNMGoogleAvailable in EU

Parametric head model with sliders for identity (253 controls), expression (383 controls), head rotation and eye direction, outputting a dense 3D head mesh including eyes, teeth and tongue. Released under Apache 2.

LULucidaByteDanceUnverified

Open-source background remover that handles glass, transparency, text, shadows and hair better than competitors shown. MIT license, under 900MB, with a free Hugging Face demo that took about 6 seconds.

MOMotion4MotionUnverified

Transfers motion between videos with different characters and proportions, without skeletons or ControlNet, by tracking points into a motion flow map and injecting it into a video model. Only a paper is released so far.

VMvidIQ MCPvidIQAvailable in EU

Sponsored segment: MCP connecting Claude, ChatGPT, Codex and other agents to Higgsfield to generate images, videos and ads.

B2Bonsai 27BPrismMLUnverified

Compressed versions of Qwen 3.6 27B: ternary at 5.9GB and 1-bit at 3.9GB, small enough for an iPhone 17 Pro. Ternary retains about 95% of baseline quality (80.5% vs 85% average benchmark score).

KKKimi K3Moonshot AIUnverified

Top open-weights model matching frontier models on many benchmarks. New leaderboard results: 2nd on debate, beats Opus 4.8 on KernelBench CUDA, 69% on DeepSWE, but lower on LiveBench; also designed a chip autonomously over 48 hours.

Full-sized T-800 humanoid robots fought in the first humanoid MMA event in Shenzhen, teleoperated, with many missed strikes.

Hoverboard-style eVTOL aircraft that flew the robots to the arena; designed to carry a robot or humans.

Hybrid robot with a humanoid upper body and four-wheeled legs for industrial settings like nuclear plants and mines, reportedly carrying up to 210 kg.

Offline robot control model for household tasks, reportedly 99% zero-shot success in unseen homes and 99.1% on laundry folding; learning vacuuming, organizing and zipping.

GEGenCeptionGoogle DeepMindUnverified

Turns a video model into a unified vision system predicting depth, normals, camera pose, segmentation, DensePose and 3D keypoints, outperforming specialist models. Only a paper is released so far.

Internal automated red-teaming model trained via self-play to find prompt-injection attacks, reaching about 84% attack success versus 13% for humans. Its attacks were used to train GPT 5.6, which had six times fewer failures than 5.5; it stays internal.

WSWan Streamer v0.3AlibabaUnverified

Real-time interactive character video that now lets the character interact with its surroundings and follow commands, using a world plus event-stream design. Only a paper is released.

ISInkling SmallThinking MachinesAvailable in EU

Open-weights multimodal MoE of almost 1T parameters (41B active) with 1M context, trained on 45T tokens; also a 276B Inkling small preview. Creator found it behind GLM 5.2 and Kimi K3 on Artificial Analysis and slower, but liked its audio and vision.

Family of embedding models (8B and 1B) for retrieval, RAG and coding agents, outperforming Qwen 3 and Gemma embeddings and cost-efficient. Multilingual, released on Hugging Face.