Quanta AI Labs

42 / Oct 13 – Oct 19, 2025 · 21 RELEASES

AI News Week 42, 2025

Ring-1T brings 1T-param reasoning. World Labs RTFM generates real-time interactive 3D worlds. NVIDIA DGX Spark arrives for developers. DiT360 masters panoramic generation.

Language Models (1)

  • Ring-1T

    InclusionAI · Reasoning

    1T-parameter deep reasoning model. "Flow State Leads to Sudden Enlightenment" with Icepop stable RL and ASystem trillion-scale training framework. Continuously evolving reasoning capabilities.

Vision & Image (5)

  • DeepSomatic

    Google Research · Visual Understanding

    AI model identifying genetic variants in tumors. Trained to spot somatic mutations in cancer cells, advancing precision oncology and tumor analysis.

  • TAG

    Research · Text-to-Image

    Tangential Amplifying Guidance for hallucination-resistant diffusion sampling. Theoretical framework with practical improvements for more faithful image generation.

  • DiT360

    Insta360 Research · Text-to-Image

    High-fidelity panoramic image generation via hybrid training with perspective and panoramic data. Circular padding, yaw loss, and cube loss ensure geometric fidelity.

  • Puffin

    Research · Visual Understanding

    Thinking with Camera — unified multimodal model for camera-centric understanding and generation. Puffin-4M dataset enables world exploration, spatial imagination, and photographic guidance. ICLR 2026.

  • MediaPipe Face Detector

    Google AI Edge · Visual Understanding

    On-device face detection via MediaPipe with BlazeFace models (short-range and full-range). Optimized for edge deployment with minimal latency.

Video & Animation (4)

  • RTFM

    World Labs · Real-time Video

    Real-Time Frame Model — generative world model that creates interactive video frames in real-time. Posed frames as spatial memory enable persistent, explorable 3D worlds.

  • DreamOmni2

    Research · Video Editing

    Multimodal instruction-based editing and generation model for unified image and video creation from natural language instructions.

  • MVP4D

    Research · Text-to-Video

    Multi-View Portrait Video Diffusion for animatable 4D avatars from a single reference image. Uses FlowFace tracking and FLAME 3DMM for precise control.

  • StreamingVLM

    MIT Han Lab · Video Understanding

    Real-time understanding for infinite video streams. Vision-language model processing continuous video with constant memory, enabling always-on AI perception.

Audio & Speech (2)

  • Vexa

    Vexa AI · Speech-to-Text

    Open-source meeting transcription API for Google Meet, Teams, and Zoom. Auto-join bots, real-time WebSocket transcripts, and MCP server for AI agents.

  • ebook2audiobook

    Community · Text-to-Speech

    Generate audiobooks from e-books with voice cloning across 1158+ languages. Supports custom TTS models and fine-tuning.

3D & Spatial (1)

  • UP2You

    Research · 3D Reconstruction

    Fast reconstruction of yourself from unconstrained photo collections. Novel paradigm for personal 3D avatar creation from casual photos.

Agents & Automation (4)

  • D2E

    WoRV AI · Robotics

    Scaling vision-action pretraining on desktop data for transfer to embodied AI. Generalist Inverse Dynamics Model enables YouTube-to-robot transfer learning.

  • PhysHSI

    Research · Robotics

    Real-world generalizable and natural humanoid-scene interaction system. Physics-based framework for natural human-scene interactions.

  • Open SWE

    LangChain · Coding Agents

    Open-source asynchronous coding agent from LangChain for automated software engineering tasks and issue resolution.

  • OpenCode

    SST / Anomaly · Coding Agents

    Open-source coding agent with desktop app (beta). 769+ contributors, 720+ releases. Supports multiple agents and extensible architecture.

Developer Tools (4)

  • DGX Spark

    NVIDIA · AI Infrastructure

    Personal AI supercomputer arriving for developers worldwide. Desktop-class hardware bringing datacenter-grade AI compute to individual researchers and developers.

  • OpenGlass

    Based Hardware · Edge Deployment

    Turn any glasses into AI-powered smart glasses. Open-source hardware and software for wearable AI (now evolved into Omi project).

  • RAG-Anything

    HKUDS · Data Integration

    All-in-One RAG framework with multimodal content understanding, knowledge graph indexing, and modality-aware retrieval. Handles documents, images, tables, and charts.

  • bRAG-langchain

    bragAI · Data Integration

    Complete RAG tutorial with LangChain covering setup, multi-query, routing, indexing, advanced retrieval, and reranking with Cohere and Pinecone.

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