24 / Jun 9 – 15, 2025 · 11 RELEASES

AI News Week 24, 2025

Week 24 2025 AI releases.

Language Models (2)

Qwen3-235B-A22B-Instruct
Alibaba Qwen · Text Generation
A 235 billion parameter instruction-tuned large language model from Qwen, representing an advanced frontier model in the dense LLM category with enhanced instruction-following capabilities.
Qwen3-Coder
Alibaba Qwen Team · Code Generation
480B-parameter MoE model with 35B active parameters trained on 7.5T tokens, supporting 256K native context and 1M with extrapolation, achieving state-of-the-art performance on agentic coding tasks through scaled Code RL and long-horizon Agent RL training. Open-sourced with Qwen Code CLI tool and Claude Code integration for real-world software engineering tasks.

DAViD
Microsoft Mixed Reality & AI Lab Cambridge · Visual Understanding
Data-efficient vision model trained exclusively on 300K high-fidelity synthetic images that achieves state-of-the-art accuracy on dense prediction tasks (depth estimation, surface normal estimation, soft foreground segmentation) while requiring significantly lower training and inference costs than foundational models. Demonstrates that procedurally-generated synthetic data with perfect labels can match or exceed real-data training performance while enabling explicit control over data diversity and fairness.

SeC - Concept-Driven Video Object Segmentation
SJTU, CUHK, Shanghai AI Lab · Video Understanding
SeC shifts video object segmentation from appearance matching to progressive concept construction using Large Vision-Language Models, achieving 11.8-point improvement over SAM 2.1. Introduces SeCVOS benchmark with 160 annotated videos for evaluating semantic understanding under complex visual variations and occlusions.

Diffuman4D
Zhejiang University & Ant Research · 3D Generation
Spatio-temporal diffusion model that generates 4D-consistent multi-view videos from sparse-view input videos, then reconstructs high-fidelity 4D Gaussian Splatting models for free-viewpoint human performance rendering. Uses skeleton-Plücker conditioning and sliding iterative denoising to synthesize novel views of humans in motion with complex clothing.
Ultra3D
Nanyang Technological University, Math Magic, Tsinghua University, Beijing Normal University, Westlake University · 3D Generation
Ultra3D is an efficient 3D generation framework that uses Part Attention, a geometry-aware localized attention mechanism, to generate high-fidelity 3D meshes at 1024 resolution with up to 6.7× speedup. The two-stage approach leverages VecSet representation for coarse layout generation followed by per-voxel latent refinement within semantically consistent part regions.

RunAgent
runagent-dev · Agent Framework
An agent framework for building and running autonomous agents with task execution capabilities. Provides infrastructure for agent orchestration and deployment.
OpenAGI Workers
OpenAGI / AI Planet · Agent Framework
Worker component system for OpenAGI framework enabling distributed task execution and agent coordination. Provides infrastructure for deploying and managing autonomous agent workers at scale.
Enhanced Agentic RAG
Uber · Agent Framework
Uber presents improvements to agentic retrieval-augmented generation systems that enhance agent reasoning and information retrieval capabilities. The system enables more sophisticated agent decision-making by integrating dynamic retrieval mechanisms with language model reasoning.

Unsloth RL Guide
Unsloth · AI Infrastructure
Documentation guide for implementing reinforcement learning with Unsloth's optimized framework, enabling efficient RL training on consumer hardware through memory-optimized techniques.
Firecrawl Observer
Mendable AI · Web Scraping
Web scraping and crawling tool that extracts structured data from websites using AI-powered content understanding. Designed for developers to convert web content into machine-readable formats for LLM applications.

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