Diffusion Technology Achieves 10x Faster Text Generation

Inception Labs has introduced the Mercury diffusion language model family, a new approach to speed up text generation. Unlike traditional sequential (autoregressive) language models, Mercury uses diffusion technology, promising significant improvements in speed and efficiency. While currently focused on code generation, this technology could transform the entire field of text generation.

How Diffusion Models Work

Diffusion models gradually recover clean, meaningful information from noisy data. The process has two steps:

  • Forward Process: Noise is added step by step to real data until it becomes random noise.

  • Reverse Process: The model learns to remove the noise, eventually producing high-quality data.

Based on principles from non-equilibrium thermodynamics, diffusion models offer advantages like more stable training, easier parallel processing, and flexible design. This helps them outperform traditional GANs or autoregressive models in tasks like generation.

Inception Labs’ Mercury Models

Unlike traditional models (which generate text left-to-right), Mercury uses a “coarse-to-fine” approach. Starting with pure noise, it refines the output over multiple steps.

Its main application today is code generation. Mercury Coder provides an interactive preview of generated code, improving developers’ workflows by showing how random characters evolve into functional code. The model can generate thousands of tokens per second—up to 10 times faster than traditional methods. Mercury is also available in downloadable versions, making it easy for businesses to integrate into their systems.

Potential Impact of Diffusion Technology

  • Speed & Efficiency: Runs on standard GPUs, speeding up development cycles and application response times.

  • Lower Cost: Works with existing infrastructure, reducing the need for specialized hardware.

  • New Research Opportunities: Combining diffusion and autoregressive models could advance tasks requiring structured logic, like coding or math problem-solving. 

Share this post
China Wants to Buy 100,000 Nvidia Chips – But Where Will the Restricted Technology Come From?
As political and economic tensions between the United States and China continue to rise, global technological competition shows no signs of slowing. Nvidia, one of the world’s most prominent chip manufacturers, is working on new strategies to maintain its presence in China, despite increasing geopolitical constraints. The company is seeking room to maneuver not only in business, but also through diplomacy.
Phase Transition Observed in Language Model Learning
What happens inside the "mind" of artificial intelligence when it learns to understand language? How does it move from simply following the order of words to grasping their meaning? A recently published study offers a theoretical perspective on these internal processes and identifies a transformation that resembles a physical phase transition.
How AI is Helping to Reduce Carbon Emissions in the Cement Industry
One industry alone is responsible for around eight percent of global carbon emissions: cement production. That’s more than the entire aviation sector emits worldwide. As the world increasingly relies on concrete for housing, infrastructure, and industrial facilities, cement manufacturing remains highly energy-intensive and a major source of pollution. A research team at the Paul Scherrer Institute (PSI) in Switzerland is aiming to change this—by using artificial intelligence to develop new, more environmentally friendly cement formulas.
Where is Artificial Intelligence Really Today?
The development of artificial intelligence has produced spectacular and often impressive results in recent years. Systems like ChatGPT can generate natural-sounding language, solve problems, and in many tasks, even surpass human performance. However, a growing number of prominent researchers and technology leaders — including John Carmack and François Chollet — caution that these achievements don’t necessarily indicate that artificial general intelligence (AGI) is just around the corner. Behind the impressive performances, new types of challenges and limitations are emerging that go far beyond raw capability.
SEAL: The Harbinger of Self-Taught Artificial Intelligence
For years, the dominant belief was that human instruction—through data, labels, fine-tuning, and carefully designed interventions—was the key to advancing artificial intelligence. Today, however, a new paradigm is taking shape. In a recent breakthrough, researchers at MIT introduced SEAL (Self-Adapting Language Models), a system that allows language models to teach themselves. This is not only a technological milestone—it also raises a fundamental question: what role will humans play in the training of intelligent systems in the future?
All it takes is a photo and a voice recording – Alibaba's new artificial intelligence creates a full-body avatar from them
A single voice recording and a photo are enough to create lifelike, full-body virtual characters with facial expressions and emotions – without a studio, actor, or green screen. Alibaba's latest development, an open-source artificial intelligence model called OmniAvatar, promises to do just that. Although the technology is still evolving, it is already worth paying attention to what it enables – and what new questions it raises.