Quanta AI Labs

43 / Oct 20 – Oct 26, 2025 · 22 RELEASES

AI News Week 43, 2025

MiniMax-M2 claims #1 open-source globally. HoloCine rivals Sora 2 for cinematic narratives. Stable Video Infinity earns ICLR Oral. Apple releases Pico-Banana-400K editing dataset. Google announces Quantum Echoes breakthrough.

Language Models (2)

  • MiniMax-M2

    MiniMax · Code Generation

    #1 open-source model globally (Artificial Analysis). 230B MoE with 10B active params built for max coding and agentic workflows. Excels at multi-file edits, tool chaining, and BrowseComp tasks.

  • ChatGPT Atlas

    OpenAI · Reasoning

    New ChatGPT feature from OpenAI for enhanced knowledge exploration and interactive research.

Vision & Image (5)

  • DeepSeek-OCR

    DeepSeek · OCR & Document

    Contexts Optical Compression model investigating the role of vision encoders from an LLM-centric viewpoint. Supports vLLM and Transformers inference for document understanding.

  • Google Earth AI

    Google · Visual Understanding

    New geospatial reasoning, deeper insights in Google Earth, and Earth AI on Google Cloud. Enables environmental monitoring, disaster response, and urban planning with AI.

  • DyPE

    Research · Text-to-Image

    Dynamic Position Extrapolation for ultra high-resolution diffusion up to 4K×4K. Plug-and-play with step-aware scaling, no finetuning required.

  • Inpaint4Drag

    Visual AI Lab · Image Editing

    Drag-based image editing via bidirectional warping + inpainting. Real-time warping preview in 0.01s, inpainting in 0.3s. 600× faster than DragDiffusion. ICCV 2025.

  • Chandra

    Datalab · OCR & Document

    OCR model handling complex tables, forms, and handwriting with full layout preservation. CLI and vLLM server support with configurable processing.

Video & Animation (6)

  • Ditto / Editto

    Research · Video Editing

    Ditto-1M: 1 million high-fidelity video editing examples built with 12,000+ GPU-days. Editto model achieves SOTA instruction-based video editing with global and local transformations.

  • HoloCine

    Research · Text-to-Video

    Holistic cinematic multi-shot long-form video generation. Window Cross-Attention localizes prompts per shot while Sparse Inter-Shot Attention enables minute-scale coherent narratives. Rivals Sora 2.

  • Stable Video Infinity

    VITA@EPFL · Text-to-Video

    ICLR 2026 Oral. Infinite-length video generation with Error Recycling. SVI 2.0 Pro supports Wan 2.2 with ComfyUI integration for film-quality long-form generation.

  • KREA Realtime Video

    KREA · Real-time Video

    Real-time video-to-video and text-to-video generation model. Works with diffusers and custom inference code for live interactive video transformation.

  • UltraGen

    SJTU · Text-to-Video

    High-resolution video generation using hierarchical attention for improved spatial fidelity and temporal consistency in synthesized videos.

  • FlashVSR

    Research · Real-time Video

    Real-time diffusion-based streaming video super-resolution. Enables high-quality upscaling with minimal latency for live video applications.

3D & Spatial (1)

  • Nano3D

    Research · 3D Reconstruction

    Training-free approach for efficient 3D editing without masks. Enables precise 3D structure modifications from minimal inputs.

Agents & Automation (3)

  • VISTA

    Google / NUS · Agent Orchestration

    Test-time self-improving video generation agent. Multi-agent system with visual, audio, and contextual critique loops. Up to 60% pairwise win rate over baselines; 66.4% human preference.

  • AgenticSeek

    Community · Web Agents

    Fully local Manus AI alternative. No APIs, runs autonomously browsing the web, writing code, and thinking locally. Privacy-focused with Docker support.

  • SoftMimic

    Research · Robotics

    Learning compliant whole-body control from examples. Teaches robots natural motion by mimicking human demonstrations with soft compliance.

Developer Tools (2)

  • OpenEnv

    Meta PyTorch · AI Infrastructure

    Agentic execution environments interface library for RL post-training. Supports torchforge, TRL, Unsloth, SkyRL, ART, and Oumi across 63+ contributors.

  • TOON

    Community · Prompt Engineering

    Token-Oriented Object Notation — compact, schema-aware JSON for LLM prompts. Saves tokens while maintaining retrieval accuracy. TypeScript SDK with VS Code support.

Research & Safety (3)

  • Quantum Echoes

    Google · Benchmark

    Breakthrough quantum algorithm on the Willow chip achieving verifiable quantum advantage. A major step toward real-world quantum computing applications.

  • Consumer Intent AI

    Research · Research Paper

    LLMs reproduce human purchase intent via Semantic Similarity Elicitation of Likert Ratings. Full backend API, web app, and benchmarking system.

  • Pico-Banana-400K

    Apple · Dataset

    400K large-scale dataset for text-guided image editing from Apple. Two-stage pipeline with automatic quality evaluation. Supports single-turn and multi-turn editing with failure cases for preference learning.

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