#transformer

'Step-by-step Tinygrad guide that builds tensors, attention, transformer blocks and a mini-GPT from scratch, with full code and experiments on training and kernel fusion.'

'IBM launches Granite 4.0 Nano: eight compact models (350M and ~1B) in hybrid SSM and transformer variants, built for local and edge inference with enterprise-grade governance and open licensing.'

'MapAnything is a unified transformer model that directly outputs factored metric 3D scene geometry from images and optional sensor inputs, achieving state-of-the-art results across multiple reconstruction tasks and datasets.'

'Step-by-step guide to create a Gemini-powered PaperQA2 agent that can query, compare and cite multiple research papers with example code and best practices.'

BAAI presents OmniGen2, a cutting-edge unified model for multimodal AI that excels in text-to-image generation, image editing, and contextual consistency, setting new open-source performance benchmarks.

Fudan University researchers have developed Lorsa, a sparse attention mechanism that disentangles atomic attention units hidden in transformer superposition, enhancing interpretability of large language models.