Скачать или смотреть Inception Mercury 2: First Thinking Diffusion Model. Frontier in LLM Speed. Reasoning Mercury dLLM

Today, we’re talking about a model that breaks the typewriter. We’re diving into Mercury 2, the first commercial-scale Diffusion-Based Language Model (dLLM) with production-grade reasoning. Instead of writing left-to-right, Mercury 2 generates the entire response as a 'noisy' sketch and then refines it all at once—reaching speeds of over 1,000 tokens per second on NVIDIA Blackwell GPUs."

In this episode, we’re unpacking why this matters for more than just speed:

The 'Beyond' Metric: Why functional correctness isn't enough anymore, and how Mercury is benchmarking code based on how fast it actually runs.

Parallel Token Prediction: How SSD-LM and the Mercury Coder family use 'coarse-to-fine' refinement to maximize hardware intensity.

Direct Preference Optimization (DPO): The training secret that aligns these diffusion models with high-performance coding standards.

We aren't just looking at a faster chatbot; we're looking at the architecture that makes AI agents and real-time voice assistants sustainable. It’s time to move beyond the sequence. Let's talk about the evolution of diffusion-based reasoning