Quanta AI Labs

13 / Mar 23 – 29, 2026 · 25 RELEASES

AI News Week 13, 2026

Week 13 2026 AI releases.

Language Models (1)

  • Gemini 3.1 Flash

    Google · Text Generation

    Fast multimodal LLM optimized for developer use cases with real-time capabilities. Designed for low-latency inference across text, image, and video inputs with streaming support.

Vision & Image (2)

  • RealRestorer

    Yang et al. (SUSTech, StepFun, SIAT-CAS) · Image Restoration

    Generalizable real-world image restoration framework leveraging large-scale image editing models. Introduces RealIR-Bench benchmark for evaluating restoration quality across diverse degradation types.

  • MegaFlow

    ETH Zurich, Microsoft, University of Tübingen · Visual Understanding

    Zero-shot optical flow model achieving state-of-the-art performance on large displacement motion estimation and point tracking using frozen DINO features, alternating frame/global attention, and recurrent refinement. Demonstrates lowest EPE on Sintel benchmark and superior zero-shot tracking on TAP-Vid.

Video & Animation (6)

  • Muse Studio

    benjiyaya · Video Editing

    Local-first workspace for planning and iterating on video narratives. Combines kanban-style scene organization, multi-LLM support (OpenAI, Claude, Ollama), ComfyUI integration for image/video gen, and an agentic Video Editor Agent.

  • Matrix-Game 3.0

    Skywork AI · Real-time Video

    Memory-augmented interactive world model achieving 720p real-time video generation at 40 FPS with a 5B model, using error-aware self-correction and camera-aware memory retrieval for minute-long temporal consistency across action-conditioned sequences.

  • daVinci-MagiHuman

    GAIR-NLP · Text-to-Video

    Unified 15B-parameter single-stream Transformer for talking-head video generation. Supports 7 languages, generates 256p-1080p video. 80% win rate vs Ovi 1.1 in human evaluation. Apache 2.0 licensed.

  • ActionPlan

    University of Tübingen, Max Planck Institute for Informatics · Text-to-Video

    A unified motion diffusion model for text-to-motion generation that enables real-time streaming (5.25× faster) and high-quality offline generation through per-frame action planning. Achieves 18% FID improvement over prior methods and supports zero-shot motion editing and in-betweening.

  • LumosX

    Zhejiang University, DAMO Academy (Alibaba Group), Hupan Lab, National University of Singapore · Text-to-Video

    A diffusion-based framework for personalized multi-subject video generation that enforces face-attribute alignment through Relational Self-Attention and Cross-Attention mechanisms. Achieves identity-consistent video generation with fine-grained subject-attribute control using subject-specific relational priors extracted via MLLMs.

  • RealMaster

    Meta Reality Labs, Tel Aviv University · Video Editing

    Converts synthetic rendered video into photorealistic video using video diffusion models with IC-LoRA, preserving 3D geometry and dynamics from 3D engines while achieving global semantic transformation of materials and lighting. Demonstrates cross-simulator generalization and dynamic weather effect editing.

Audio & Speech (2)

  • PrismAudio

    Huawei / FunAudioLLM · Audio Processing

    Video-to-audio framework using decomposed Chain-of-Thought reasoning across semantic, temporal, aesthetic, and spatial dimensions. Introduces Fast-GRPO training and AudioCanvas benchmark. Accepted to ICLR 2026.

  • Transcribe

    Cohere · Speech-to-Text

    Audio transcription service leveraging Cohere's language models for accurate speech-to-text conversion with support for multiple languages and audio formats.

3D & Spatial (3)

  • RetimeGS

    HKUST, Netflix Eyeline Labs, HKUST(GZ) · NeRF & Gaussian

    A 4D Gaussian Splatting method for continuous-time reconstruction of dynamic scenes at arbitrary timestamps, addressing temporal aliasing through optical flow-guided initialization, triple-rendering supervision, and spline trajectory modeling to enable ghost-free interpolation under large inter-frame motion.

  • World Reconstruction From Inconsistent Views

    Lukas Höllein, Matthias Nießner (TUM) · 3D Reconstruction

    Reconstructs 3D worlds from video diffusion models by resolving frame inconsistencies through non-rigid alignment and iterative frame-to-model ICP, followed by 2DGS optimization to generate sharp, explorable 3D environments from inconsistent multi-view sequences.

  • LagerNVS

    University of Oxford, Meta AI · 3D Reconstruction

    Real-time neural novel view synthesis using 3D-biased encoder-decoder architecture with latent geometry representation. Achieves real-time rendering without explicit 3D representations by leveraging 3D priors through architecture design and pre-training with VGGT.

Agents & Automation (3)

  • Reply.io

    Reply.io · Agent Framework

    AI-powered sales engagement platform automating outreach across email, LinkedIn, calls, and SMS. Features AI SDR agents for lead generation, multichannel sequences, and CRM integrations.

  • WorldAgents

    Technical University of Munich (Ziya Erkoç, Angela Dai, Matthias Nießner) · Agent Framework

    Multi-agent framework that synthesizes 3D-consistent worlds from 2D foundation models using a VLM-based director, image generator, and two-step verifier, demonstrating that 2D models possess implicit 3D world understanding. Outputs explorable 3D Gaussian Splatting scenes from iterative view synthesis.

  • CUA-Suite

    ServiceNow, Mila, University of Waterloo, Université de Montréal, McGill University, University of Oxford, National University of Singapore · Desktop Agents

    A unified ecosystem for training and evaluating desktop computer-use agents, comprising 55 hours of continuous 30 fps expert video (6M frames), 56K densely annotated screenshots (5M+ UI elements), and expert-designed tasks across 87 professional applications with multi-layered reasoning annotations (~497 words per step).

Developer Tools (4)

  • RuView

    ruvnet · Edge Deployment

    Edge AI system using WiFi signals for camera-free human sensing. Detects presence, 17-keypoint pose, and vital signs through walls up to 5m. Runs on ESP32 mesh (~$8-54/deployment) with 60+ WASM edge modules.

  • SportRadar Developer API

    SportRadar · Data Integration

    REST API providing access to real-time sports data, statistics, and odds across multiple sports. Enables developers to integrate live sports information into applications.

  • Dynamic VRAM in ComfyUI

    Comfy · AI Infrastructure

    Memory optimization feature for ComfyUI that dynamically manages VRAM allocation to reduce memory overhead on local GPU inference, enabling larger model deployment on consumer hardware.

  • TurboQuant

    Google · AI Infrastructure

    Extreme compression technique for AI models achieving significant parameter reduction while maintaining performance. Enables efficient deployment of compressed models with minimal accuracy loss.

Research & Safety (4)

  • Something Big Is Happening

    Matt Shumer · Survey & Overview

    Analysis of AI capability acceleration citing GPT-5.3 Codex and Opus 4.6. Warns the gap between public perception and current AI reality is dangerous, with major white-collar disruption expected within 1-5 years.

  • Tribe V2 Brain Predictive Foundation Model

    Meta · Research Paper

    A predictive foundation model designed for brain imaging analysis and neural activity prediction. Demonstrates capabilities in processing neuroimaging data with improved accuracy over prior versions.

  • ARC-AGI-3

    ARC Prize Foundation · Benchmark

    Interactive reasoning benchmark measuring human-like intelligence in AI agents through novel environment exploration, goal acquisition, world model building, and continuous learning. Evaluates skill-acquisition efficiency, long-horizon planning, and experience-driven adaptation rather than static puzzle-solving.

  • Pulse of Motion

    Texas A&M University · Benchmark

    A method for measuring and correcting physical frame rate (PhyFPS) in AI-generated videos by analyzing visual dynamics. Detects chronometric hallucinations where motion speeds don't align with nominal frame rates, with user studies confirming that PhyFPS-corrected videos show significantly improved perceived naturalness.

Published by Quanta AI Labs, an AI agency in India. Browse every week in AI News Weekly.