Samsung’s Tiny AI Crushes Gemini 2.5 on ARC-AGI While IBM’s WebGPU Model Parses Docs Locally
AI Research & Breakthroughs
Samsung’s Tiny Recursive Models Outperform Larger AI on ARC-AGI with 0.01% Parameters
Samsung SAIL Montréal introduced the Hierarchical Reasoning Model (HRM) and Tiny Recursive Model (TRM), which use small neural networks to solve complex tasks like Sudoku and ARC-AGI while outperforming models like Gemini 2.5 Pro and DeepSeek with 10,000x fewer parameters. The research highlights a novel approach to efficient AI reasoning.
- AI 10000x smaller than Gemini 2.5 pro and deepseek beat them both in arc agi 1 and 2
- Samsung Paper Reveals a Recursive Technique that Beats Gemini 2.5 Pro on ARC-AGI with 0.01% of the Parameters!
New AI Models & Releases
Liquid AI Releases LFM2-8B-A1B: 3-4B Quality at 1.7B Speed for Edge Devices
Liquid AI’s LFM2-8B-A1B is a hybrid Mixture-of-Experts (MoE) model optimized for on-device AI, delivering performance comparable to 3-4B dense models while running faster than Qwen3-1.7B. It is designed for high-end phones, tablets, and laptops, setting a new benchmark in efficiency.
IBM’s Granite Docling: Local, WebGPU-Powered Document Parsing in the Browser
IBM launched Granite Docling, a 258M-parameter VLM for document conversion that runs 100% locally in the browser via WebGPU, ensuring privacy for sensitive documents. The demo showcases state-of-the-art parsing without server dependency.
AI Tools & Developer Integrations
Atlassian Launches Rovo Dev: Full SDLC AI Agent for Jira, GitHub, and IDEs
Atlassian’s Rovo Dev is now generally available, offering a context-aware AI agent integrated across Jira, CLI, IDEs, GitHub, and Bitbucket to streamline the software development lifecycle (SDLC) with automated assistance.
Developer Builds GPT-5 System with Live Market Data Integration
A custom system connects GPT-5 to real-time market data using FastAPI, OpenAI API, Next.js, and TradingView charts, enabling AI-driven analysis with live financial inputs.