ByteDance Seed
Seed Diffusion Preview is a large-scale, code-focused language model that uses discrete-state diffusion to generate code non-sequentially, achieving dramatically faster inference without sacrificing quality by decoupling generation from the token-by-token bottleneck of autoregressive models. It combines a two-stage curriculum, mask-based corruption followed by edit-based augmentation, to robustly train a standard dense Transformer, striking a balance between speed and accuracy and avoiding shortcuts like carry-over unmasking to preserve principled density estimation. The model delivers an inference speed of 2,146 tokens/sec on H20 GPUs, outperforming contemporary diffusion baselines while matching or exceeding their accuracy on standard code benchmarks, including editing tasks, thereby establishing a new speed-quality Pareto frontier and demonstrating discrete diffusion’s practical viability for real-world code generation.
Learn more
Mercury Coder
Mercury, the latest innovation from Inception Labs, is the first commercial-scale diffusion large language model (dLLM), offering a 10x speed increase and significantly lower costs compared to traditional autoregressive models. Built for high-performance reasoning, coding, and structured text generation, Mercury processes over 1000 tokens per second on NVIDIA H100 GPUs, making it one of the fastest LLMs available. Unlike conventional models that generate text one token at a time, Mercury refines responses using a coarse-to-fine diffusion approach, improving accuracy and reducing hallucinations. With Mercury Coder, a specialized coding model, developers can experience cutting-edge AI-driven code generation with superior speed and efficiency.
Learn more
DiffusionGemma
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.
Learn more
Mercury Edit 2
Mercury Edit 2 is part of Inception Labs’ Mercury family of AI models, designed to perform high-speed reasoning, coding, and editing tasks using a fundamentally different architecture from traditional large language models. It builds on Mercury 2, a diffusion-based reasoning model that generates and refines entire outputs in parallel rather than producing text token by token, enabling significantly faster performance and more responsive editing workflows. Instead of acting like a sequential “typewriter,” the system behaves more like an editor, starting with a rough draft and iteratively improving it across multiple tokens at once, which allows for real-time interaction and rapid iteration in tasks such as code editing, content generation, and agent-based workflows. This architecture delivers throughput of up to around 1,000 tokens per second, making it several times faster than conventional models while maintaining competitive reasoning quality across benchmarks.
Learn more
