Alvaro Cintas
Ant Group just built a diffusion model that performs surgery on its own output 🤯 LLaDA2.2-flash edits mid-generation and still keeps block-parallel speed. It catches its own mistakes, deletes them, and patches in fixes, all while running as a real agent that plans and calls tools. - Beats the top autoregressive model on τ²-Bench and MCP-Atlas - 1.64× faster throughput - Native 128K context 100% Open Source.