Quanta AI Labs

13 / Mar 24 – 30, 2025 · 8 RELEASES

AI News Week 13, 2025

Week 13 2025 AI releases.

Vision & Image (2)

  • HiDream-I1-nf4

    hykilpikonna · Text-to-Image

    Quantized version of HiDream-I1 model in nf4 format, likely a compressed variant of an image or multimodal generation model optimized for resource-constrained inference.

  • UNO: Multi-Subject Image Generation Framework

    ByteDance · Text-to-Image

    UNO is a universal framework that scales from single-subject to multi-subject image generation through progressive cross-modal alignment and universal rotary position embedding (UnoPE), enabling high-consistency controllable generation trained iteratively on synthesized multi-subject data.

Video & Animation (1)

  • Pusa-VidGen

    Yaofang Liu · Text-to-Video

    A video generation model that appears to focus on efficient video synthesis capabilities. The project is hosted on GitHub as an open-source implementation.

3D & Spatial (1)

  • HoloPart

    University of Hong Kong & VAST · 3D Reconstruction

    HoloPart introduces 3D part amodal segmentation using a diffusion-based model (PartComp) that decomposes 3D shapes into complete, semantically meaningful parts even when occluded. The method combines existing segmentation techniques with a novel architecture featuring local and global attention mechanisms to infer hidden geometry while maintaining shape consistency.

Agents & Automation (1)

  • AgentDesk Browser Tools

    AgentDesk · Web Agents

    Browser automation extension enabling AI agents to interact with web interfaces through installation and control mechanisms. Provides programmatic access to browsing capabilities for agent-based workflows.

Developer Tools (3)

  • Geospatial Reasoning

    Google Research · AI Infrastructure

    Google's geospatial AI platform integrating Gemini with foundation models for satellite imagery, mapping, and real-world problem-solving across flood forecasting, wildfire detection, weather prediction, and urban development through natural language queries and multi-modal reasoning.

  • Mistral Small 3.1 Fine-tuning Support

    Unsloth · AI Infrastructure

    Unsloth now supports fine-tuning of Mistral Small 3.1 (24B multimodal model with 128K context), achieving 1.8x faster training, 70% VRAM reduction, and requiring only 18GB minimum VRAM. Multiple quantized versions (2-8 bit GGUFs, dynamic 4-bit, 16-bit) are available on Hugging Face.

  • python-a2a

    Manoj Desai · AI Infrastructure

    Python library for Account-to-Account (A2A) payment integration, enabling direct bank transfers between accounts via APIs.

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