Quanta AI Labs

32 / Aug 4 – 10, 2025 · 10 RELEASES

AI News Week 32, 2025

Week 32 2025 AI releases.

Language Models (2)

  • Kimi-K2-Instruct-0905

    Moonshot AI · Text Generation

    An instruction-tuned variant of the Kimi-K2 large language model optimized for following user directives and conversational tasks. Released as an open-weight model on Hugging Face for research and deployment purposes.

  • LongCat-Flash-Chat

    Meituan · Text Generation

    A chat system optimized for long context processing with flash attention mechanisms, enabling efficient handling of extended conversation histories and documents.

Video & Animation (1)

  • TalkingHead

    met4citizen · Text-to-Video

    An open-source application that generates talking head videos from text or audio input, combining speech synthesis with facial animation to create realistic video avatars.

Audio & Speech (1)

  • AudioStory

    Tencent ARC · Text-to-Speech

    AudioStory is a system for generating narrative audio content that combines speech synthesis with story generation capabilities. It enables the creation of audio stories through AI-driven composition and voice synthesis.

3D & Spatial (2)

  • ReconViaGen

    CUHKSZ (Chinese University of Hong Kong, Shenzhen) · 3D Reconstruction

    A novel framework that integrates 3D reconstruction priors with diffusion-based generative models to achieve accurate multi-view 3D object reconstruction with improved consistency and completeness. Addresses key limitations of pure reconstruction methods (incomplete results) and generation-based methods (image inconsistency) through enhanced cross-view feature extraction and controllable denoising.

  • LSD-3D: Large-Scale 3D Driving Scene Generation

    Princeton University, Mercedes-Benz, Torc Robotics · 3D Generation

    A method for generating large-scale 3D driving scenes with geometry grounding by combining coarse geometric layouts, environment maps, and Gaussian splatting, optimized through diffusion-based distillation sampling to enable causally-consistent novel view synthesis with explicit 3D geometry.

Agents & Automation (1)

  • Oasis 2.0

    Decart AI · Agent Framework

    A real-time AI experience for Minecraft that dynamically transforms the game world as players interact with it. Installable via Modrinth or CurseForge as a mod with live demo available.

Developer Tools (2)

  • Embedding GeMMA 300M

    Google · AI Infrastructure

    A 300M parameter text embedding model designed for efficient semantic search and retrieval tasks. Optimized for generating high-quality vector representations with reduced computational overhead.

  • SkyPilot

    SkyPilot Organization · AI Infrastructure

    An open-source framework for running LLMs and batch jobs on any cloud (AWS, GCP, Azure) with minimal code changes and automatic resource optimization. Abstracts away cloud-specific infrastructure complexity through a unified Python interface.

Research & Safety (1)

  • Artificial Analysis Text-to-Image Arena

    Artificial Analysis · Benchmark

    Benchmarking platform for evaluating and comparing text-to-image generation models through comparative ranking. Provides performance metrics across different generative image models.

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