ByteDance Seed

DiffusionGemma is an experimental open model that explores text diffusion, an exceptionally fast approach to text generation. Released under an Apache 2.0 license, this 26B Mixture of Experts (MoE) model moves beyond the sequential token-by-token processing of typical autoregressive Large Language Models (LLMs). Instead, it generates entire blocks of text simultaneously, delivering up to 4x faster text generation on GPUs. Built on the intelligence-per-parameter of the Gemma 4 family and Gemini Diffusion research, DiffusionGemma integrates a novel diffusion head designed to maximize generation speed. It is designed for researchers and developers exploring speed-critical, interactive local workflows such as in-line editing, rapid iteration, and non-linear text structures. By shifting the decode bottleneck from memory bandwidth to compute, it can generate more than 1,000 tokens per second on a single NVIDIA H100 and more than 700 tokens per second on an NVIDIA GeForce RTX 5090.

Celeris-1 is a low-latency, general-purpose language model platform and a diffusion model designed to deliver frontier-level intelligence at dramatically higher speed. Instead of generating one token at a time like traditional autoregressive models, Celeris uses a diffusion-based inference architecture that enables parallel generation and response times measured in milliseconds. On its published MMLU-Pro benchmark, Celeris-1 reaches 75.9 accuracy with a 158 ms median response time and 1,664 output tokens per second, placing it within a few points of frontier models while running more than 10x faster. The model is exposed through an OpenAI-compatible API, so developers can point existing SDKs and clients at Celeris with minimal code changes. Streaming is enabled for interactive applications, with responses as low as 24 ms and no buffering or batch delay.