peterindia.netDiffusion Language Models
Language models that generate text by refining a whole sequence in parallel instead of predicting one token at a time: commercial dLLMs, open-weight models, the research behind them and the tools that speed them up.
Inception's family of commercial diffusion language models, presented as the first commercial-scale dLLMs and aimed at very fast generation.
Best for: Fast, low-latency text and code generation.
Visit Mercury (Inception) →Developer documentation for calling Inception's Mercury models through an API.
Best for: Trying a diffusion LLM from your own code.
Visit Inception Platform docs →OpenRouter's listing for Inception's Mercury model, with pricing and a unified API.
Best for: Testing Mercury alongside other models.
Visit Mercury on OpenRouter →Google DeepMind's page for Gemini Diffusion, an experimental model that generates text by refining noise rather than token by token.
Best for: Following Google's work on text diffusion.
Visit Gemini Diffusion →ByteDance Seed's paper on a large-scale diffusion language model with high-speed inference.
Best for: Understanding an industrial dLLM design.
Visit Seed Diffusion →Open-source masked diffusion language model released with the paper "Large Language Diffusion Models", presented as an alternative to autoregressive LLMs.
Best for: The reference open diffusion LLM.
Visit LLaDA →A 7B-parameter open diffusion language model from the HKU NLP group, with a project blog explaining its training.
Best for: Experimenting with a mid-sized open dLLM.
Visit Dream 7B →Apple research code for masked diffusion models applied to code generation.
Best for: Code generation with diffusion.
Visit DiffuCoder →Synergy of Diffusion and AutoRegression: diffusion language models released in sizes from 1.7B to 30B parameters.
Best for: Combining diffusion with autoregressive training.
Visit SDAR →Open diffusion language model for code, releasing pretraining, evaluation and inference code and checkpoints.
Best for: Studying a fully open code dLLM.
Visit Open-dLLM →Open multimodal large diffusion language models for text and images, with reinforcement-learning-based tuning.
Best for: Multimodal diffusion research.
Visit MMaDA →A large diffusion language model for multimodal understanding.
Best for: Vision-language tasks with diffusion.
Visit LaViDa →Early paper showing that continuous diffusion can control text generation, "Diffusion-LM Improves Controllable Text Generation".
Best for: The origin of text diffusion for control.
Visit Diffusion-LM →"Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution", a score-based approach to discrete diffusion for language.
Best for: Foundations of discrete diffusion.
Visit SEDD →Code for "Simple and Effective Masked Diffusion Language Models", a widely used masked-diffusion recipe.
Best for: A clear baseline for masked diffusion.
Visit MDLM →Code for "Block Diffusion", which interpolates between autoregressive and diffusion language models.
Best for: Mixing block-wise generation with diffusion.
Visit Block Diffusion (BD3-LMs) →Official code for an ICLR 2025 paper on scaling masked diffusion models on text.
Best for: Scaling-law style comparisons.
Visit Scaling up Masked Diffusion Models on Text →Code for an ICLR 2025 paper on discrete diffusion for complex reasoning and planning.
Best for: Reasoning and planning with diffusion.
Visit Beyond Autoregression →The LLaDA paper, which trains a diffusion language model from scratch and compares it with autoregressive models.
Best for: Reading the key open dLLM paper.
Visit Large Language Diffusion Models (paper) →"Mercury: Ultra-Fast Language Models Based on Diffusion", Inception's technical report.
Best for: How a commercial dLLM is described.
Visit Mercury (paper) →NVIDIA's training-free method that accelerates diffusion LLMs with KV caching and parallel decoding.
Best for: Speeding up existing dLLMs.
Visit Fast-dLLM →Repository for a survey paper on diffusion language models, with a curated list of papers.
Best for: A map of the research field.
Visit A Survey on Diffusion Language Models →Hugging Face community article introducing diffusion language models.
Best for: A readable first introduction.
Visit Diffusion Language Models: The New Paradigm →Match your requirement to a starting point, then compare vendors.
| A fast hosted dLLM to try | Mercury through the Inception API or OpenRouter |
|---|---|
| Run an open model locally | LLaDA or Dream 7B |
| Code generation research | DiffuCoder or Open-dLLM |
| Image plus text with diffusion | MMaDA or LaViDa |
| Learn how they work | Diffusion-LM, SEDD, MDLM and the survey |
| Speed up decoding | Fast-dLLM and block-wise approaches |
Neighbouring directories and hubs.
Quick answers.
A diffusion language model (dLLM) generates text by starting from a masked or noisy sequence and refining it over several steps, usually predicting many tokens in parallel, instead of producing one token after another.
Autoregressive models write left to right and cannot revise earlier tokens. Diffusion models work on the whole sequence, can fill in or edit text in any order, and can generate several tokens per step.
The main promise is faster generation and more controllable, editable output. Limits include a smaller ecosystem, less mature serving tools and, for many tasks, quality that still trails the best autoregressive models. Test on your own workload.
A few are offered commercially, such as Inception's Mercury models, while others, including Gemini Diffusion, are experimental. Most open models are research releases.
Not at present. They are an active research direction and an alternative approach, and hybrid designs such as block diffusion mix the two.
Descriptions summarise each project's own pages and papers, checked on 5 October 2026. Diffusion language models are a fast-moving research area, so confirm availability and licences on the linked pages.