Full body waifus, AI dreams, realtime AI music, open-source Gemini Omni: AI NEWS
AI Search showed 25 AI tools in this video on . 8 are verified available in the EU. Each row below jumps to the exact second. Part of week 2026-23.
Tools shown in this video, with timestamps
Open-source unified video editing model that edits videos via text prompts or image/video references (object removal, background swap, style change, reference insertion). Based on Wan 2.2 with high/low branches, ~84GB per branch; ComfyUI integration underway.
3D reconstruction model that turns multiple images into 3D Gaussian splats with camera poses by reusing one transformer block repeatedly. At 117M parameters it matches Depth Anything 3, which is about 10x larger; code released.
Panoramic Geometry Reconstruction model that predicts depth and surface normals from a single 360° panorama by adapting a 3D foundation model to cube faces. Model (~6GB) and datasets released.
Open-source real-time music generator controllable live via MIDI, audio and text. Latency drops to ~200ms from 3s, 2.4B parameters, runs on a MacBook, aimed at DAW plugin use.
Live demoChatGPT memory upgrade that synthesizes memory from past conversations in the background to keep context fresh rather than relying on stale saved notes. Improves accuracy and recall over time; rolling out to free users with larger memory for paid users.
Markerless multi-person motion capture from multi-view video, recovering pose, shape, and hands of each person even in close contact. Trained on the Mamma-Syn synthetic dataset of 2.5M crops; code released.
Closed-source, paid image model that first lays out bounding boxes in structured form, giving composition control and separate editable regions. Ranks #2 on Arena text-to-image behind GPT Image 2 and ahead of Nano Banana 2.
Image model with open weights, strong in typography and posters, using JSON prompts with bounding-box layout and layered outputs, including transparent images. ComfyUI workflow available; heavily censored; top open model on Arena, ranked ninth overall.
Open Apache 2 encoder-free multimodal model taking text, images and audio directly, runs offline in 16GB memory. Performs close to the roughly twice-larger 24B variant and shows audio transcription and voice editing.
Live demoMultimodal agentic model for long-horizon tasks, shown coding an app autonomously for 11 hours and testing it. Beats DeepSeek V4 Pro and GLM 5.1 on average; available only via Cloud Studio and API.
Open-source physical AI foundation model combining text, image, video, audio and action to generate synthetic data for robots and autonomous vehicles. Nano and Super variants, with models, training scripts and datasets released.
Superchip for laptops and small desktops combining a Blackwell GPU, 20-core CPU, up to 1 petaflop FP4 and up to 128GB unified memory for running local AI models. Laptops planned for fall.
Decomposes an image into transparent editable layers, trained with reinforcement learning using a VLM scorer instead of paired data. Beats Qwen Image Layered; only a paper so far.
Industrial-grade all-weather humanoid robot with dust and water protection, shown sprinting, flipping a circuit breaker switch and traversing hills and stairs.
Teaser demo of a realistic full-body emotional humanoid robot couple, planned to debut later this month.
Upcoming open-source multimodal model with 1M context, sparse attention and strong agentic coding, beating GPT-5.5 on SWE-Bench Pro. Leads open models on Artificial Analysis but lags GLM 5.1 on Arena; weights due in 1-2 weeks.
Microsoft quantum chip developed with help from its Discovery AI system, with a new material stack claimed 1000x more reliable than the previous qubits. Microsoft expects a scalable quantum computer by 2029.
Zero-shot voice cloning TTS built on F5-TTS that clones a voice from a few seconds of audio. Released with training scripts; checkpoint is 10GB.
Live demoReal-time streaming avatar generator from a prompt and transcript, with motion control and consistent videos over 5 minutes. Runs in real time on a single H100; only a paper released.
Live demoReal-time generative world model for autonomous driving simulation producing multi-camera driving videos from frames, text, maps and trajectories, with weather changes and editing. Code released.
Open 550B mixture-of-experts model with 55B active, hybrid Mamba-transformer, NVFP4 and multi-token prediction, 1M context. Claimed 5x faster inference and about 30% lower cost than other open models; weights are 350GB.
Open TTS model with inline tags to control emotion, style, speed, pitch, pauses and sound effects. Tops other open TTS models in comparisons; 9.3GB on Hugging Face.
Live demoVideo generation framework from Baidu's ERNIE team that adds natively aligned audio to Wan 2.2, producing 720p video with synchronized speech. Reportedly beats LTX 2.3 despite a far smaller size; open-source, 25GB with a 7GB FP8 version.
Live demoMicrosoft's 1T-parameter mixture-of-experts thinking model with 35B active, outperforming Sonnet 4.6 on several benchmarks but behind top models. In private preview in Microsoft Foundry.
Closed, paid image generation and editing model available in Microsoft Foundry. Ranks above Nano Banana models on Arena but behind GPT Image 2.
