33 / Aug 11 – 17, 2025 · 12 RELEASES

AI News Week 33, 2025

Week 33 2025 AI releases.

Language Models (2)

ERNIE-4.5-21B-A3B-Thinking
Baidu · Reasoning
A 21B parameter language model from Baidu featuring extended thinking capabilities and chain-of-thought reasoning for complex problem-solving tasks.
MedGemma
Google DeepMind · Text Generation
A specialized medical language model based on Gemma that demonstrates improved performance on biomedical question answering and medical reasoning tasks through domain-specific training.

BiRefNet
MatteoKartoon · Visual Understanding
A machine learning model for biological reference network analysis and segmentation tasks. Designed to process and segment biological imagery with deep learning techniques.
UniVG-R1
Tsinghua University & Alibaba Group · Visual Understanding
A reasoning-guided multimodal LLM for universal visual grounding that uses GRPO reinforcement learning with cold-start initialization and a difficulty-aware weight adjustment strategy, achieving 23.4% average improvement on zero-shot reasoning-grounding benchmarks and 9.1% improvement on MIG-Bench.

Deep-Live-Cam
hacksider · Video Inpainting
A deepfake video synthesis tool that generates realistic real-time facial animations by mapping source video to target faces using deep learning techniques. Enables live video transformation with neural face-swapping and expression transfer capabilities.
Vizard API
Vizard · Video Editing
API for automating social-ready short-form video generation from long-form content using AI-powered clipping and transcription. Supports 30+ languages, customizable branding templates, and integrates seamlessly into existing platforms with 99.99% uptime.

Stable Audio 2.5
Stability AI · Music Generation
Enterprise-grade audio generation model designed for production-scale sound synthesis and audio creation workflows. Optimized for high-quality, controllable audio output in professional applications.

FE2E: Dense Geometry Estimator
Beijing Jiaotong University, Alibaba Group, Chongqing University of Posts and Telecommunications, Nanyang Technological University · Depth Estimation
FE2E is a DiT-based foundation model that adapts image editing models rather than generative models for monocular dense geometry prediction, achieving 35% performance gains on depth estimation and joint depth-normal prediction with only 71K training images through reformulated flow matching loss and consistent velocity training.
WinT3R
Shanghai AI Lab, University of Science and Technology of China, Zhejiang University · 3D Reconstruction
A feed-forward 3D reconstruction model that performs online camera pose estimation and point map generation using a sliding window mechanism and camera token pool for efficient streaming reconstruction. Achieves state-of-the-art quality and speed on diverse datasets while processing image streams in real-time.

NLWeb
Microsoft · Web Agents
NLWeb integrates conversational AI interfaces directly into web browsers, enabling natural language interaction with web content and services. The system leverages language models to provide a more accessible and intuitive web browsing experience.

Pathway
Pathway · Data Integration
Open-source data processing framework for building real-time AI applications with streaming data pipelines and LLM integrations.
LLM App
Pathway · AI Infrastructure
Open-source framework for building LLM-powered applications with real-time data processing and streaming capabilities. Provides abstractions for connecting language models to live data sources and building production-ready AI applications.

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