Diffusion language models challenge GPT with a 892-token speed record

# Diffusion Language Models Challenge GPT: Speed Record of 892 Tokens per Second

Diffusion models are rewriting the rules: 100 billion parameters, 892 tokens per second

The language model industry has received an unexpected challenge. Researchers have demonstrated that diffusion approaches to text generation, long considered slow and inefficient, are not only able to compete with classical architectures like GPT, but can surpass them in speed. A diffusion language model with 100 billion parameters achieved a record generation speed of 892 tokens per second — a figure that calls into question established views on how modern large language models should work.

This achievement is particularly significant because diffusion methods have remained on the periphery in the context of text tasks for many years. While in computer vision diffusion models gained authority and revolutionized image synthesis, autoregressive paradigm prevailed in text processing — the same one on which ChatGPT and its competitors are built. Autoregressive models predict the next token based on all previous ones, which requires sequential passes through the neural network and slows down the process.

Diffusion language models work on fundamentally different principles. Instead of generating text word by word, they start with noisy data and gradually refine the output through several denoising stages. The paradox: with such an approach, seemingly requiring more computational operations, the new 100-billion model showed a speed of 892 tokens per second. This is approximately twice as fast as typical performance metrics of modern autoregressive models of similar size. The technical breakthrough lies in optimizing the denoising algorithm and network architecture, which allows parallel processing of multiple positions in text, rather than waiting for one token prediction to complete before moving to the next.

The significance of this result extends far beyond simply setting a speed record. Successful scaling of the diffusion model to 100 billion parameters proves that this approach is not an engineering dead end. If diffusion models can operate with such performance, they open new paths for optimization. Manufacturers can reduce latency, improve server throughput, and decrease energy consumption — critical factors in the era of cloud computing costs.

For the industry, this means the future of language models is not necessarily tied to autoregressive architecture. OpenAI, Google DeepMind, and other labs have invested enormous resources in optimizing the autoregressive approach, but the emergence of a competitive alternative may force strategy reconsideration. Companies that have invested in researching diffusion methods gain tangible advantage. For end users, this could mean faster responses from AI assistants, cheaper APIs, and more energy-efficient local models.

However, caution should be exercised in interpreting the results. Token generation speed is far from the only criterion for model quality. Text quality, ability to handle long-term dependencies, and logical consistency are also important. It remains to be understood whether the diffusion approach can match autoregressive models in terms of content richness and accuracy of responses under equal computational resources.

This event symbolizes a transitional moment in the AI industry, when the dominant paradigm begins to feel competition. If diffusion models confirm their viability on other parameters as well, we may witness true architectural diversity in mainstream AI, each with its own strengths.