Quantum breakthrough, Gemini 4 Argon, MOSS Image, KLING 4, OpenAI spy, Muse Wonder: AI NEWS
AI Search showed 25 AI tools in this video on . 9 are verified available in the EU. Each row below jumps to the exact second. Part of week 2026-42.
Tools shown in this video, with timestamps
Open-source text-to-speech model supporting meta tags to control emotion and delivery. Ranked number one among open models on the Artificial Analysis leaderboard; 35B total parameters with 3B active, reportedly runs on 8 GB VRAM.
Live demoWorld model that creates playable real-time 3D games with consistent mechanics. A game engine handles physics and rules while the AI renders detailed visuals; includes a coding agent and runs at 83 fps 480p on one consumer GPU.
Live demoAnimates 3D characters by combining skeleton motion with a flexible outer-surface representation, handling clothing, fur and animals. Code released.
Live demoTurns a few multi-angle images of an object into a full 3D model, combining reconstruction of visible parts with generation of missing geometry. Preserves more detail than compared methods; code and models released.
Live demoGoogle demonstrated magic state cultivation on its Willow quantum processor, cutting errors about 40x and reaching 99.99% fidelity. It could reduce hardware requirements for quantum computers, though larger-scale demos are still needed.
Memory technique for video generators that selectively retains and retrieves details from earlier frames so objects and scenes stay consistent after being hidden. Two models released, text-to-video based on Minimax H3 and image-to-video based on Ling World Infinity.
Live demoGoogle's latest flagship model, released to Pro/Ultra and paid API users. Strong benchmark results (top on Vals Index, Blueprint Bench, medical reasoning, near-saturating ARC AGI 3 with a harness) and up to 1M output tokens.
Multi-player world model that simulates several players in the same environment with synchronized per-player camera views. Trained with four players, it scaled to 10 without extra training; hierarchical communication cuts compute by over 86%. Code released.
Live demoNew Kling video model with more dynamic motion, stereo audio, 10-bit HDR, up to 15 references and multi-keyframe control. The creator found it close to but not as good as SeaArt in initial testing.
Live demoImage generation and editing model producing 2K images in under 8 seconds, versus 30+ seconds for Nano Banana. Edits preserve untouched pixels and support up to six references; currently paid-only with open-sourcing hinted.
Small 5.6B open image generation/editing model with 4K support and bounding box control, available in ComfyUI under a non-commercial community license. The creator's tests showed bad faces and anatomy, so he passes on it.
Live demoOpen 5B text-to-image model under Apache 2, relatively unfiltered, good at anime, game scenes and text rendering, with bounding-box layout control. Released as a base model and an SFT model; no editing.
Meta's new frontier model for agentic and professional work, succeeding Muse Spark. It is fast and hallucinates less, but is expensive per task and ranks around top five or lower on several independent benchmarks.
Open-weights 816B MoE model (42B active, 1M context) claimed to be the most capable US open-weight model. It ranks below other open models like GLM 5.3 on the intelligence index, costs more, and has a restrictive license with required attribution and registration.
System that keeps long generated videos consistent by reconstructing a 3D environment as memory, so rooms stay the same when revisited. Works with different video models; code released.
Live demo3D reconstruction model that turns multi-view photos into explorable 3D scenes, including large driving-footage environments, with cleaner results than other methods. About 6.6 GB, code and models released.
Live demoWorld model based on Minimax H3 that converts rough game-engine output into realistic visuals in real time, using two-step generation. Needs eight GPUs and over 200 GB of files; model and code released.
Live demoWorld model that transforms a rough 3D environment into realistic video one segment at a time, producing a segment in 5.1 seconds on one H200. More efficient than Nvidia's earlier method; code and training scripts released.
Live demoOpen-source full-stack humanoid robot platform including hardware designs, electronics, firmware, assembly docs and software. Configurations range from 24 to 42 degrees of freedom with a 500 Hz real-time control system.
Live demoSystem that lets two humanoid robots play doubles table tennis, combining movement and racket skills with a planner coordinating both robots. Demo shows long rallies.
Live demoStepFun's flagship open-weights model for autonomous coding and reasoning, 600B MoE with 27B active and 1M context. Among the best open models, slightly behind GLM 5.3 and Mimo 2.6, at about $1 per task via API; 650 GB.
Open multimodal model (120B, 12B active) that understands images, video and documents, with over 600 tokens per second. Not the most intelligent open model, but training recipe is released; about 75 GB.
Open model with 560B total and 25B active parameters, performing close to leading open models like GLM 5.3, Kimi and DeepSeek 4.1. About 574 GB.
Small open model with 24B parameters and 2B active, designed for high-end consumer devices. Outperforms similar-sized models like Qwen 3.6 and Gemma 4 on average, is fast, 48 GB, and mostly Apache 2.0.
Language modeling approach that generates text from coarse concepts to fine details, producing tokens in parallel within each stage. Reported up to 21x faster than other diffusion language models and 31x faster on summarization; code released.
