New AI waifus, new Deepseek, realtime worlds, Happy Shrimp, tiny TTS: AI NEWS

EVEvokeUnverified

Open-source 14B real-time interactive world video model taking an image plus joystick input, generating video in three steps; supports text-prompted events. Apache 2 licensed, final model ~57 GB.

4D4DAnyoneUnverified

Turns a single video of a person into a 4D Gaussian splat viewable from any angle, via skeleton extraction and multi-view generation. More detailed than ReCamMaster and TrajectoryCrafter; model is 12 GB.

SUSenseNova U1.5 8BSenseTimeAvailable in EU

Open-source image generator and editor that generates native 4K end-to-end in pixel space without a VAE. Handles posters, infographics and natural-language editing; ~50 GB, Apache 2.

BVBernini v2ByteDanceAvailable in EU

Open-source omnimodal video editor that edits videos via natural language, including adding or removing objects, camera changes and reference insertion. The model is ~180 GB, so the creator thinks it may be dead on arrival.

O1Ornith 1.5Available in EU

Open model family (9B dense, 35B MoE, 397B MoE) trained with a self-improvement loop where the system proposes problems, builds scaffolds and generates RL data. The largest beats GLM 5.2 and nears Opus 4.8 on agentic benchmarks; GGUFs released.

ATAudio8 TTS 0.1BEdge0Available in EU

A 0.1B-parameter multilingual text-to-speech model with voice cloning from a few seconds of audio. Total size is 1.7 GB, so it runs on consumer devices.

GEGeoWeaverUnverified

Reconstructs long videos into consistent 3D scenes by chunking, predicting depth and camera pose, then stitching chunks to reduce drift. Lowest average error on benchmarks; only a paper is released so far.

QVQwen Video EditAlibabaUnverified

Video editing from text prompts that runs Qwen Image Edit through a Wan video generation workflow. Code and training script released; 41 GB total, and the creator sees little point given Minimax H3.

Data collection glove with fingertip and palm tactile sensors (0.1 N sensitivity), a wrist camera and angular encoders for gathering hand data to train robots.

Transformer robot that switches between biped humanoid, quadruped, drone-carried form, and a coming wheeled form. Creator doubts practicality due to more failure points; not released.

Elf-themed bionic humanoid robot shown at the World Robot Conference; now with an articulated body and expanded degrees of freedom. The creator calls Headform's robots the most realistic.

Rideable quadruped robot horse for rough terrain, with 300 kg payload and 40 km range at about $40,000 (X1); a wheel-legged XS version costs about $53,000.

USUnitree SupermanUnitreeUnverified

Previewed humanoid that does a ~2 m standing high jump and sprints at ~12.7 m/s, beating human records. Developed in about 3 months and has trouble stopping.

System that transfers real tennis players' moves onto a Unitree G1 humanoid, enabling autonomous rallies and serves.

DVDeepSeek V4 Flash Vision ExperimentalDeepSeekAvailable in EU

Adds image, video and document understanding to V4 Flash while matching its text performance. It matches Opus 4.8 on several agentic coding benchmarks; API only for now.

HSHappy ShrimpUnverified

Music generator from the same lab as Happy Horse; takes style descriptions and lyrics or instrumental mode. The creator calls it among the best music generators and notes it is free for now.

VMvidIQ MCPvidIQAvailable in EU

Open-sourced agentic connector letting AI agents like Codex connect to a local ComfyUI install, understand workflows, models and hardware, and build and run workflows from natural language.

G1Gen 1.5Generalist AIUnverified

Robot foundation model that can attempt a task from a single 3-12 second demonstration. It reaches 59% success across 10 tasks one-shot and 83% with about 5 minutes of extra data.

Agentic harness that raised Claude Opus 5 from 30% to 100% on the public ARC-AGI-3 set, completing all 183 levels across 25 environments. Creator notes the public set may inflate the score.