07 / Feb 10 – 16, 2025 · 11 RELEASES

AI News Week 7, 2025

Week 7 2025 AI releases.

Language Models (2)

QwQ-32B
Qwen Team · Reasoning
A 32B parameter open-weight model achieving DeepSeek-R1 performance through reinforcement learning scaling, using outcome-based rewards on math/coding tasks followed by general capability training. Includes integrated agent capabilities and tool usage.
Babel
Alibaba DAMO Academy · Multilingual
Open-source multilingual LLM covering 25 languages spoken by 90% of global population, with 9B and 83B variants using layer extension techniques. Babel-83B achieves state-of-the-art performance among open multilingual models, comparable to GPT-4o on certain tasks.

Diffusion Self-Distillation (DSD)
Stanford (Cai, Chan, Zhang, Guibas, Wu, Wetzstein) · Image-to-Image
A method that leverages pre-trained text-to-image diffusion models to generate paired training data through self-distillation, then fine-tunes them into text+image-to-image models for identity-preserving generation and other conditional tasks without requiring test-time optimization.
Aya Vision
Cohere · Visual Understanding
Multimodal AI model that combines vision and language capabilities for understanding and reasoning about images across multiple languages.
YOLOE
THU-MIG (Tsinghua University) · Visual Understanding
An efficient object detection model extending YOLOv8 with optimized architecture for real-time detection tasks. Provides improved speed-accuracy tradeoffs through architectural innovations and training enhancements.
YOLOE
jameslahm · Visual Understanding
Object detection model based on YOLO architecture optimized for efficient inference. Available as a Hugging Face Space with GPU acceleration.

GEN3C
NVIDIA Toronto AI Lab · Text-to-Video
A generative video model that enables precise camera control and temporal 3D consistency by using a 3D cache of point clouds derived from depth predictions, achieving state-of-the-art results in novel view synthesis from single images, sparse views, and dynamic videos.

NotaGen
Central Conservatory of Music, Tsinghua University, et al. · Music Generation
NotaGen is a symbolic music generation model pre-trained on 1.6M musical pieces and fine-tuned on 9K classical compositions using LLM paradigms, employing a novel CLaMP-DPO reinforcement learning method to enhance musicality and controllability without human annotations. The model uses hierarchical GPT-2 decoders with interleaved ABC notation and demonstrates superior performance to baselines in generating high-quality classical sheet music.
DiffRhythm
ASLP Lab, Northwestern Polytechnical University · Music Generation
A latent diffusion-based model that generates full-length songs (up to 4m45s) with synchronized vocals and accompaniment in ~10 seconds using only lyrics and style prompts. Non-autoregressive architecture enables fast inference while maintaining a simple, scalable design without complex multi-stage pipelines.

OpenManus
mannaandpoem · Robotics
OpenManus is an open-source framework for robotic manipulation and hand control, enabling dexterous task execution through unified motor control interfaces.

HunyuanVideo Training
spacepxl · AI Infrastructure
Open-source training implementation for Hunyuan Video, a text-to-video generation model. Includes training code, datasets, and configurations for reproducing video synthesis capabilities.

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