Quanta AI Labs

44 / Oct 27 – Nov 2, 2025 · 20 RELEASES

AI News Week 44, 2025

Kimi-Linear redefines efficient attention with 6x decoding throughput. AMD Nitro-E trains image gen in 1.5 days. WorldGrow generates infinite 3D worlds. MiniMax Music 2.0 launches pro-grade AI composition.

Language Models (2)

  • Kimi-Linear

    Moonshot AI · MoE Architecture

    Hybrid linear attention architecture with Kimi Delta Attention (KDA). 48B-A3B MoE reduces KV cache by 75% and boosts decoding throughput 6x for 1M-token contexts while outperforming full attention.

  • Aardvark

    OpenAI · Text Generation

    New model release from OpenAI pushing the frontier of language understanding and generation capabilities.

Vision & Image (4)

  • Nitro-E

    AMD AGI · Text-to-Image

    Ultra-efficient 304M text-to-image diffusion model. Trains in 1.5 days on a single 8×MI300X node, generates 512px images at 18.8 samples/s. Distilled version reaches 39.3 samples/s. Runs on consumer iGPU in 0.16s.

  • SVG Latent Diffusion

    Research · Text-to-Image

    Latent diffusion model without variational autoencoder, exploring novel approaches to image generation with reduced architectural complexity.

  • GRAG Image Editing

    Research · Image Editing

    Group-Relative Attention Guidance for precise image editing control in DiT models. Integrates in 4 lines of code, enabling continuous fine-grained editing intensity control without tuning.

  • Pomelli

    Google Labs · Text-to-Image

    Experimental Google Labs project exploring novel AI-powered creative tools and interactive experiences.

Video & Animation (2)

  • LongCat-Video

    Meituan · Text-to-Video

    Long-form video generation framework from Meituan enabling temporally coherent, high-quality video synthesis over extended durations.

  • MoCha

    Orange 3DV Team · Video Editing

    End-to-end video character replacement without structural guidance. Requires only a single first-frame mask and reference images. Handles occlusion, rare poses, and complex illumination with temporal consistency.

Audio & Speech (3)

  • Foley Control

    Stability AI · Audio Processing

    Video-guided Foley sound generation using frozen models. Connects V-JEPA2 video embeddings to Stable Audio Open DiT via compact cross-attention bridge. Achieves competitive sync with far fewer trainable params.

  • Udio × Universal Music Group

    Udio · Music Generation

    Partnership between Udio and UMG enabling AI-powered music creation with licensed artist voices and styles. Features remix, reimagine, and mashup tools for fan creativity with artist permissions.

  • MiniMax Music 2.0

    MiniMax · Music Generation

    Next-generation AI music generation with lifelike vocals, diverse singing styles, precise instrument control, and professional-grade audio quality. Available via API and web app.

3D & Spatial (2)

  • WorldGrow

    Research · 3D Generation

    Hierarchical framework for generating infinite, unbounded 3D worlds via block-by-block synthesis. Uses 3D block inpainting and coarse-to-fine strategy for photorealistic, structurally consistent scene generation.

  • iGGT

    Research · 3D Reconstruction

    Instance-Grounded Geometry Transformer for semantic 3D reconstruction. End-to-end framework with InsScene-15K dataset supporting instance-level scene understanding from synthetic and real-world data.

Agents & Automation (4)

  • Game-TARS

    ByteDance Seed · Desktop Agents

    Extension of UI-TARS GUI agent model for game interaction. Next-gen native GUI agent designed for human-like perception and seamless interface interaction.

  • Lemon AI

    Hexdo · Coding Agents

    First full-stack open-source self-evolving general AI agent. Fully local alternative to Manus/Genspark with integrated Code Interpreter VM sandbox, HTML editor, and self-learning from conversations.

  • DWService Agent

    DWService · Desktop Agents

    Cross-platform remote access agent for Linux, Mac, and Windows. Open-source agent enabling secure remote machine management and control.

  • Agent Lightning

    Microsoft · Agent Framework

    Train ANY AI agent with RL using zero code changes. Supports LangChain, OpenAI SDK, AutoGen, CrewAI, and more. Features selective multi-agent optimization with RL, prompt optimization, and SFT.

Developer Tools (2)

  • GenAI Toolbox

    Google · AI Infrastructure

    Open-source MCP server for databases. Integrates AI agents with databases in <10 lines of code with connection pooling, auth, and OpenTelemetry observability. Supports natural language to SQL.

  • DS/ML Projects Collection

    Community · AI Infrastructure

    Curated collection of Data Science and Machine Learning projects with complete source code for learning and reference.

Research & Safety (1)

  • Efficient Attention (arXiv)

    Research · Research Paper

    Research paper on efficient attention mechanisms for transformer models, exploring sparsity and computational optimization techniques.

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