Quanta AI Labs

39 / Sep 22 – 28, 2025 · 11 RELEASES

AI News Week 39, 2025

Week 39 2025 AI releases.

Language Models (3)

  • Hunyuan-A13B

    Tencent · Small Language Models

    A 13 billion parameter language model from Tencent designed for efficient inference and deployment. Part of the Hunyuan family of models with focus on practical applications.

  • LongWriter-Zero-32B

    THU-KEG · Text Generation

    A 32B parameter language model designed for long-form content generation with zero-shot capabilities. Optimized for extended text generation tasks without task-specific fine-tuning.

  • HY-MT1.5-7B

    Tencent · Multilingual

    A 7B parameter model from Tencent focusing on machine translation capabilities. Appears to be an updated version of their translation model with improved performance.

Vision & Image (1)

  • Z-Image

    Tongyi-MAI · Text-to-Image

    Image generation or manipulation model from Tongyi-MAI research team. Specific capabilities unavailable due to truncated page content.

Video & Animation (1)

  • Lumiere Video Stylization

    Google Research · Video Editing

    AI system for applying artistic styles to videos while maintaining temporal coherence and preserving motion dynamics. Enables video-to-video style transfer with physics-aware stylization across frames.

Audio & Speech (1)

  • MedASR

    Google · Speech-to-Text

    A medical-domain automatic speech recognition model optimized for healthcare transcription tasks, leveraging domain-specific training to improve accuracy on clinical terminology and medical conversations.

3D & Spatial (2)

  • DreamCube

    University of Hong Kong, Tencent, Astribot · 3D Generation

    DreamCube introduces Multi-plane Synchronization to adapt 2D diffusion models for cubemap-based 3D panorama generation, enabling RGB-D panorama synthesis, depth estimation, and consistent omnidirectional content creation from single-view inputs and text prompts.

  • Depth Anything at Any Condition (DepthAnything-AC)

    Nankai University (Boyuan Sun, Modi Jin, Bowen Yin, Qibin Hou) · Depth Estimation

    Foundation monocular depth estimation model that handles diverse environmental conditions including adverse weather, illumination variations, and sensor distortions using unsupervised consistency regularization finetuning and spatial distance constraints. Achieves zero-shot generalization across indoor/outdoor/night/adverse weather scenarios with minimal unlabeled data (540K images).

Agents & Automation (2)

  • SearchAgent-8B-hq

    aidenjhwu · Agent Framework

    An 8B parameter language model optimized for search and information retrieval tasks with high-quality outputs. Designed as a lightweight agent framework for web search and query resolution.

  • OpenManus

    OpenManus · Robotics

    Reinforcement learning framework for robotic manipulation and control tasks. Provides tools for training agents to perform dexterous hand and arm movements.

Developer Tools (1)

  • vLLM-Omni

    vLLM Project · Inference Engine

    An extension of vLLM adding multimodal capabilities for efficient inference of vision-language models. Optimizes token generation across text, image, and audio modalities with batching and memory management improvements.

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