Mercury 2: The First Reasoning Diffusion Language Model (1,000+ tokens/sec)

Quick Overview

Inception Labs released Mercury 2, the first reasoning Diffusion Large Language Model (dLLM), which achieves over 1,000 tokens per second inference speed by generating tokens in parallel using a diffusion-based approach, significantly outpacing traditional autoregressive LLMs which require sequential, left-to-right generation.

Key Points: Mercury 2 is the first commercial-grade diffusion LLM, utilizing a diffusion-based approach for language generation, enabling parallel token generation. The model achieves over 1,000 tokens per second (t/s) inference speed, demonstrated to be significantly faster than autoregressive models like ChatGPT, which completed a 75-iteration task in 75 iterations versus Mercury's 14 iterations. The diffusion approach starts with noise and iteratively refines the output in parallel over multiple passes, contrasting with autoregressive models that generate one token sequentially. Mercury 2 demonstrates competitive reasoning quality, scoring 91.1 on AME 2025 and 73.6 on GPQA, while maintaining its speed advantage. The input cost for Mercury 2 is $0.25 per 1M tokens, and the output cost is $0.75 per 1M tokens, positioning it as a cost-effective option compared to other top-tier models. Inception Labs also offers Mercury Edit, a smaller, coding-focused dLLM optimized for low-latency coding tasks.

Context: Inception Labs announced the release of Mercury 2, positioning it as a significant advancement in Large Language Models (LLMs) by applying diffusion-based techniques, traditionally used in image and video generation, to text generation. This marks a departure from the standard autoregressive method used by models like GPT-3 and LLaMA, aiming to deliver superior speed and efficiency while maintaining high reasoning quality.

Detailed Analysis

Inception Labs introduced Mercury 2, the first reasoning Diffusion Large Language Model (dLLM), which leverages a diffusion-based approach to achieve unprecedented speed in text generation, reaching over 1,000 tokens per second (t/s). This contrasts sharply with traditional autoregressive LLMs (like ChatGPT shown in a benchmark), which generate text token by token sequentially, resulting in significantly higher iteration counts for the same task (e.g., 75 iterations for autoregressive vs. 14 for Mercury 2). The speed advantage is inherent to the model architecture, allowing parallel token generation across multiple refinement passes starting from noise, rather than relying on better hardware or infrastructure optimization alone. Benchmarks show Mercury 2 maintains competitive intelligence, achieving 91.1 on AME 2025 and 73.6 on GPQA, while offering favorable pricing ($0.25 input/ $0.75 output per 1M tokens). The company also details Mercury Edit, a coding-focused dLLM, and emphasizes that their foundational technology is applicable beyond text to image and video generation. The platform provides an OpenAI-compatible API, supporting integration with existing developer tools and workflows.