Quanta AI Labs

51 / Dec 15 – Dec 21, 2025 · 53 RELEASES

AI News Week 51, 2025

GPT-5.2 and Devstral 2 advance frontier models, Gemini 2.5 TTS and GLM-4.6V push multimodal, with RealGen and Wan-Move leading video/image generation.

Language Models (4)

  • GPT-5.2

    OpenAI · Text Generation

    Latest GPT iteration with enhanced reasoning, extended context, and improved multimodal capabilities.

  • GLM-4.6V

    Zhipu AI · Text Generation

    Open-source multimodal models with native tool use, visual web search, and frontend replication capabilities.

  • Gemini 3 Flash

    Google · Text Generation

    Frontier intelligence built for speed with global availability, offering state-of-the-art performance at scale.

  • Bolmo-1B

    Allen AI · Small Language Models

    Efficient 1B parameter language model from Allen Institute for AI with open weights.

Vision & Image (8)

  • RealGen

    Research · Text-to-Image

    Photorealistic text-to-image framework using detector-guided rewards to eliminate AI artifacts.

  • Qwen-Image-i2L

    DiffSynth Studio · Image-to-Image

    Image-to-LoRA model for style transfer, enabling fine-grained control over diffusion generation.

  • WindowSeat Reflection Removal

    Huawei BayerLab · Image Restoration

    AI-powered reflection removal from window seat photos using deep learning.

  • NewBie-image

    NewBie AI Lab · Text-to-Image

    Experimental image generation model on ModelScope with advanced capabilities.

  • ChatGPT Images

    OpenAI · Text-to-Image

    New native image generation capabilities in ChatGPT with improved quality and multimodal understanding.

  • FLUX.2 Max

    Black Forest Labs · Text-to-Image

    Top-tier image generation model delivering the highest quality output with unmatched photorealism and editing capabilities.

  • Qwen-Image-Layered

    Qwen · Image Editing

    Image decomposition model enabling flexible layered editing and iterative refinement for design applications.

  • SVG-T2I

    Kling Team · Text-to-Image

    Scaling up text-to-image latent diffusion without variational autoencoder using novel SVG autoencoder.

Video & Animation (18)

  • Wan-Move

    Alibaba VILab · Text-to-Video

    NeurIPS 2025 paper on motion-controllable video generation via latent trajectory guidance with ComfyUI support.

  • One-to-All Animation

    Research · Video Editing

    Alignment-free character animation and image pose transfer for TikTok and cartoon datasets.

  • EgoEdit

    Snap Research · Video Editing

    Real-time egocentric video editing at 855ms first-frame latency for live AR interactions.

  • StereoWorld

    Research · Text-to-Video

    Geometry-aware monocular-to-stereo video generation with 11M training dataset.

  • Light-X

    Research · Text-to-Video

    Generative 4D video rendering with camera and illumination control for diverse lighting conditions.

  • OneStory

    Research · Text-to-Video

    Coherent multi-shot video generation with adaptive memory for minute-long narrative videos.

  • Saber

    Research · Text-to-Video

    Scalable zero-shot reference-to-video generation trained on video-text pairs with emergent abilities.

  • Seedance 1.5 Pro

    ByteDance Seed · Text-to-Video

    Film-grade video generation with native audio, strong cinematography, emotional storytelling, and SeedVideoBench-1.5 evaluation.

  • Ray3

    Luma Labs · Text-to-Video

    AI video generation with HDR support, character reference, keyframe editing, reasoning, and SOTA physics in Dream Machine.

  • Kling AI Global

    Kuaishou · Text-to-Video

    Global platform for Kling AI video generation models with comprehensive creative tools.

  • Wan AI Platform

    Alibaba · Text-to-Video

    AI creative platform with text-to-image, image-to-image, text-to-video, and image editing capabilities.

  • LongVie 2

    Vchitect · Text-to-Video

    Multimodal controllable ultra-long video world model for 3-5 minute generation with world-level guidance.

  • RealVideo

    Zhipu AI · Text-to-Video

    Real-time streaming conversational system powered by autoregressive diffusion video generation with Dynamic Sink RoPE.

  • LongCat-Video-Avatar

    Meituan · Text-to-Video

    Expressive audio-driven avatar model for super-realistic lip-synchronized long video generation with consistent identity.

  • MiMo V2 Flash

    Xiaomi · Text-to-Video

    Fast video generation model from Xiaomi with advanced AI capabilities.

  • SCAIL

    Research · Video Editing

    Studio-grade character animation via in-context learning of 3D-consistent pose representations with DiT architecture.

  • V-RGBX

    Research · Video Editing

    Video editing with accurate controls over intrinsic properties including relighting and multi-attribute editing.

  • PersonaLive

    GVCLab · Video Editing

    Expressive portrait image animation for live streaming with online and offline inference support.

Audio & Speech (4)

  • Gemini 2.5 TTS

    Google · Text-to-Speech

    New Flash and Pro text-to-speech models with enhanced style versatility, pacing control, and multi-speaker capabilities.

  • SAM Audio

    Meta AI · Audio Processing

    First unified multimodal model for audio separation combining Perception Encoder with audiovisual understanding.

  • Chatterbox Turbo

    Resemble AI · Text-to-Speech

    Fast open-source TTS model with multi-language support and PerTh watermarking for responsible AI.

  • PolyAI Voice Hub

    PolyAI · Voice Agents

    Voice AI resources and documentation hub for building conversational voice assistants.

3D & Spatial (6)

  • MoCA

    Research · 3D Generation

    Mixture-of-Components Attention for scalable compositional 3D generation with efficient routing.

  • MoCapAnything

    Huawei · 3D Reconstruction

    Unified 3D motion capture for arbitrary skeletons from monocular videos with cross-species retargeting.

  • TRELLIS.2

    Microsoft · 3D Generation

    Native and compact structured latents for high-resolution 3D asset generation with 4B-parameter model and PBR support.

  • TRELLIS.2-4B

    Microsoft · 3D Generation

    Hugging Face release of the 4B parameter 3D generation model with sparse compression VAE and instant mesh conversion.

  • Hunyuan 3D World

    Tencent · 3D Generation

    Tencent Hunyuan 3D generation platform supporting text and image to 3D asset creation with diffusion technology.

  • StereoSpace

    ETH Zurich · 3D Reconstruction

    Stereo depth estimation and 3D reconstruction web demo from ETH Zurich research.

Agents & Automation (7)

  • Devstral 2 & Vibe CLI

    Mistral AI · Coding Agents

    State-of-the-art open-source agentic coding models with production-grade workflows and CLI agent.

  • AutoGLM Open Source

    Zhipu AI · Agent Framework

    32 months of development now open-sourced for building autonomous agents with browser and device control.

  • A2UI

    Google · Agent Framework

    Agent-to-User Interface framework for building AI agent interactions with Lit renderer and shell client.

  • Claude Code Templates

    AITMPL · Agent Framework

    Professional templates for Anthropic Claude Code with 100+ agents, commands, settings and hooks.

  • Mino

    Mino AI · Web Agents

    AI-powered web automation that turns natural language into browser actions with real-time streaming.

  • Hindsight

    Vectorize · Agent Framework

    Agent memory system that works like human memory with retain, recall, reflect operations and learning capabilities.

  • AI Data Science Team

    Business Science · Agent Framework

    AI-powered team of agents for data science tasks including Pipeline Studio app and multi-LLM support.

Developer Tools (5)

  • TwinFlow

    Research · Inference Engine

    Self-adversarial flows for one-step generation on large models like Qwen-Image-20B without distillation.

  • TurboDiffusion

    Tsinghua ML · Inference Engine

    100-200x acceleration for video diffusion models like Wan2.1 and Wan2.2 with optimized inference.

  • Banana Slides

    Community · AI Infrastructure

    AI-native PPT generator based on Nano Banana Pro with template support, smart parsing, and natural language editing.

  • TabPFN

    Prior Labs · Inference Engine

    Foundation model for tabular data with sklearn-compatible interface and instant predictions without training.

  • Node Banana

    Community · AI Infrastructure

    Visual node editor for Nano Banana with workflow creation, connection rules, and TypeScript support.

Research & Safety (1)

  • HubSpot AI Video Guide

    HubSpot · Survey & Overview

    Comprehensive comparison guide for AI video generators covering features and use cases.

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