News
ReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior ReleaseReleasesOVHcloud unveils Olala mini, its first in-house LLMChinaA 9B model beats a 1T model on FinQA, trained with Crusoe Managed RLReleasesThe internal metaphysics of an LLMReleasesMapping Language Model Internal Representations to Physical Laws and Scientific ReasoningResearchCan China keep rogue AI in check? Meet the start-up testing the top modelsFrontierMicrosoft aims to challenge OpenAI, Anthropic with cheaper, specialized AI modelsOpen modelsMicrosoft rolls out new first-party models and updates to agentic AI continuous improvementFrontierMicrosoft unveils AI models for coding, cybersecurity as it takes on Anthropic, OpenAIFrontierGoogle is making it easier to opt out of Gemini AI model training – here’s howFrontierOpenAI Unveils New Astra AI Model After Scrapping Prior Release
EveryLLM
Seed Diffusion Preview

Seed Diffusion Preview

Fast discrete diffusion language model for code

Seed Diffusion Preview is a large language model built on discrete-state diffusion architecture. Released in August 2025, it prioritises rapid inference speed for coding tasks by generating tokens in parallel rather than sequentially.

Share on WhatsApp

  • Uses discrete-state diffusion for parallel token generation
  • Achieves 2,146 tokens per second on H20 GPUs
  • Faster than Mercury and Gemini Diffusion models
  • Focuses on code generation benchmarks
  • Released as a research preprint in August 2025

What it is

Seed Diffusion Preview is a research model designed to generate text using a diffusion process rather than standard autoregressive methods. It focuses specifically on coding applications where speed is critical. The system uses discrete states to allow parallel token generation, which avoids the delays common in sequential decoding. This approach aims to balance high performance with rapid output, making it suitable for real-time developer tools. The model was introduced in mid-2025 as part of ongoing efforts to improve inference efficiency in large language systems.

How it came about

The model was announced in August 2025 via a preprint paper on arXiv. The work was conducted by a team of researchers including Yuxuan Song and Zheng Zhang among others. It builds on recent developments in diffusion-based language models such as Mercury Coder and Gemini Diffusion. The release highlights a shift towards non-sequential generation techniques to address latency issues inherent in traditional transformer decoders. No prior versions or specific development timeline details are provided in the available material.

Inside the model

The core architecture relies on discrete-state diffusion, which differs from continuous diffusion used in image generation. This design enables the model to predict multiple tokens simultaneously instead of one by one. By avoiding sequential dependency during generation, the system achieves significant speed improvements. The training process optimises for both accuracy and inference velocity. The authors note that this parallel generation mechanism is key to reducing the time required to produce code completions while maintaining quality standards expected in professional programming environments.

What it can do

The model demonstrates strong performance on standard code evaluation benchmarks. Its primary advantage lies in inference speed, reaching 2,146 tokens per second when running on H20 graphics processing units. This rate is notably higher than comparable models like Mercury and Gemini Diffusion. The developers claim it sets a new benchmark for the trade-off between speed and quality in code generation tasks. While specific scores on individual tests are not detailed, the overall competitive standing is emphasised as a key achievement.

How to use it

Access is currently limited to research purposes through the published paper and associated project page. A demo is available online for users to test the capabilities. Interested parties can view the technical details via the arXiv repository.

Why it matters

The introduction of Seed Diffusion Preview contributes to the growing interest in diffusion models for text generation. It challenges the dominance of autoregressive transformers in coding assistants by offering a faster alternative. If the speed advantages hold in broader applications, it could influence how future code generation tools are built. The work highlights the potential of parallel decoding strategies to overcome latency bottlenecks in large-scale language models.

Questions people ask about Seed Diffusion Preview

What makes Seed Diffusion different from other models?

It uses discrete-state diffusion to generate tokens in parallel rather than sequentially. This allows for much faster inference speeds compared to traditional autoregressive models that predict one word at a time.

How fast is the model?

The model reaches an inference speed of 2,146 tokens per second when tested on H20 GPUs. This is significantly faster than contemporary alternatives like Mercury and Gemini Diffusion.

Is the model open source?

The provided material does not state that the weights are open source. It is available as a research paper on arXiv with a demo link, but full access terms are not specified.

What is it used for?

It is designed primarily for code generation tasks. The model aims to provide high-speed completions for programmers while maintaining competitive performance on standard coding benchmarks.

Good reading and sources

    Where this page comes from

    There is no English Wikipedia article on Seed Diffusion Preview yet, so the facts on this page come from the sources listed above: arxiv.org/abs/2508.02193. Some are the subject's own pages, so its claims about itself are reported as its claims. EveryLLM wrote the text in its own words, with AI assistance, and compares it with the sources to make sure no passage is copied.

    It may contain mistakes, and figures change over time. Nothing here is investment or buying advice. If something is wrong, tell us. Last refreshed 2026-11-14.

    The weekly digest

    Every Monday: the model of the day, new pages on EveryLLM, and the conferences and anniversaries of the week ahead.