Quanta AI Labs

47 / Nov 17 – Nov 23, 2025 · 15 RELEASES

AI News Week 47, 2025

Depth Anything 3, SAM 3, Grok 4.1, and Gemini 3 headline a stacked week across vision and reasoning.

Language Models (5)

  • Grok 4.1

    xAI · Reasoning

    Latest Grok release levels up coding + reasoning benchmarks with “thinking tokens” visibility.

  • GPT-5.1 Codex Max

    OpenAI · Code Generation

    Code-focused GPT variant layers long-horizon planning, multi-file edits, and integrated evals.

  • Uni-MoE v2

    Idealist Labs · MoE Architecture

    Upgraded mixture-of-experts stack mixes sparse + dense routing for efficient multimodal tasks.

  • Gemini 3

    Google · Text Generation

    Third-gen Gemini pushes unified multimodal reasoning plus faster tool-calling reliability.

  • Gemini Nano Banana Pro

    Google · Small Language Models

    New Nano variants bring on-device multimodal reasoning to Android hardware.

Vision & Image (3)

  • Segment Anything Model 3

    Meta AI · Visual Understanding

    SAM3 unifies text, points, and images in one transformer for fast class-agnostic segmentation.

  • Antigravity

    Google Labs · Text-to-Image

    Web experiment lets users sculpt physics-bending worlds powered by Gemini creative tools.

  • Part-X-MLLM

    Chunshi Wang · Visual Understanding

    Multimodal large language model learns part-level grounding for robotics and CAD.

3D & Spatial (2)

  • Depth Anything 3

    HKU & ByteDance · Depth Estimation

    Next-gen monocular depth model brings SAM-style prompts plus open-weight checkpoints for robotics and AR.

  • SAM 3D

    Meta AI · 3D Reconstruction

    Extends SAM3 to 3D assets, plugging segmentation into reconstruction and NeRF workflows.

Agents & Automation (1)

  • Proactive Hearing

    UW & Microsoft Research · Embodied AI

    Earable agents anticipate speech and environmental cues to surface proactive assistance.

Developer Tools (3)

  • Kandinsky Lab

    Kandinsky · Inference Engine

    Browser playground for Kandinsky diffusion with storyboards, reference images, and live style transfer.

  • WeatherNext 2

    Google DeepMind · AI Infrastructure

    Improved AI forecasting stack hits 10-day accuracy with faster inference for operational meteorology.

  • PhysX Anything

    PhysX Anything Team · AI Infrastructure

    Open toolkit injects physics priors into diffusion models for controllable simulation assets.

Research & Safety (1)

  • DR-TULU

    RL Research · Benchmark

    Benchmarks direct preference optimization for reinforcement-style instruction tuning.

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