2 min read

Google’s VISTA Self-Improving Video AI & AMD ROCm 7.9 Expand GPU Support

Research & Model Innovations

Google Introduces VISTA: A Self-Improving Video Generation Agent
Google Research unveiled VISTA, a multi-agent system that iteratively refines video generation through structured temporal planning, pairwise tournaments, and specialized critique agents (visual, audio, contextual). The system demonstrates improvements in visual fidelity, storytelling, and transitions by synthesizing feedback into enhanced prompts. Examples and technical details are available in the paper.


Cerebras Releases REAP-Pruned Lightweight Models (GLM4.5-Air & Qwen3-Coder-30B)
Cerebras published REAP (Routed Expert Adaptive Pruning) checkpoints for GLM4.5-Air (25% pruned) and Qwen3-Coder-30B (20% pruned), achieving minimal accuracy loss while reducing model size. The pruned models are available on Hugging Face, offering efficient alternatives for deployment.


Model Releases & Framework Updates

Llama.cpp Adds Support for InclusionAI’s Ling & Ring Models (1000B/103B/16B)
The Ling and Ring model families (including 1T, 103B, and 16B variants) from InclusionAI are now compatible with llama.cpp, enabling local inference for these large-scale models. Hugging Face hosts the model checkpoints, with flash-attention and mini versions also available.


AMD ROCm 7.9 RC1 Released with Strix Halo Support
AMD’s latest ROCm 7.9 Release Candidate 1 expands hardware compatibility to include the Strix Halo GPU, improving accessibility for AI/ML workloads. The update targets performance optimizations and broader hardware support for ROCm-based applications.


Platform & Developer Tools

Mistral AI Launches Redesigned Documentation Site
Mistral AI unveiled a revamped documentation portal with improved navigation, detailed model/endpoint guides, and clearer explanations of its platform features (e.g., agents, APIs). The update aims to enhance developer onboarding and usability.


Gelt.Dev: Affordable Multi-Agent Coding Tool Gains Traction
Gelt.Dev, a cost-effective multi-agent coding assistant, was featured as the #7 product of the day on a launch platform. The tool leverages collaborative AI agents to streamline development tasks, targeting budget-conscious users.


Industry Insights & Announcements

Google’s Gemini Team Teases Future Plans in Upcoming Podcast
Logan Kilpatrick (Google Gemini team) will join The Roo Cast podcast to discuss the roadmap for Gemini, AI advancements, and strategic directions. The episode promises insights into Google’s next steps for its flagship AI model.