A language model that generates text by denoising a whole sequence in parallel — like an image model — instead of predicting one token at a time.
Last verified · 2026-08-31 · by Moe Ameen
A diffusion language model (DLM, sometimes dLLM) is a large language model that produces text by iterative denoising rather than next-token prediction. Where an autoregressive model like GPT or Claude writes one token at a time, strictly left to right, a diffusion model starts from a fully corrupted sequence — every position masked or filled with noise — and refines the entire sequence in parallel across a series of steps until coherent text emerges. It is the same core idea behind AI image generators such as Stable Diffusion, applied to language: corrupt data into noise, then learn to reverse the process.
The defining consequence is that generation is no longer strictly ordered. On each step the model looks at the whole (still-noisy) sequence and revises every position at once, with context in both directions, rather than appending words to the end. A helpful analogy is an editor sharpening a rough draft of the entire answer in passes versus a typist producing final words one after another. The number of denoising steps is a dial — more steps generally means higher quality, fewer means faster output — and getting good text from as few steps as possible is much of what the field works on.
There are two families. Discrete diffusion corrupts text by masking or swapping actual tokens and learns to recover them, which fits language's categorical nature and is what most usable 2025–2026 models employ. Continuous diffusion instead maps tokens to embedding vectors, adds Gaussian noise as an image model would, then decodes back to tokens; it is harder to train but distills better into very-few-step samplers. The headline appeal of the whole approach is speed — parallel decoding removes the one-token-at-a-time bottleneck, and commercial diffusion LLMs have been measured at over a thousand tokens per second — traded against training efficiency and, in strict parallel sampling, some coherence between tokens generated in the same step.
Diffusion models were popularized for images in the early 2020s, and researchers quickly asked whether the same denoising idea could generate text. Early continuous attempts (Diffusion-LM and related work around 2022) proved the concept but trained far less efficiently than autoregressive baselines, and after 2023 discrete, masking-based methods took the lead because they mapped more naturally onto language's categorical structure.
The paradigm crossed into serious contention in 2025–2026. LLaDA (Large Language Diffusion with mAsking), reported in early 2025, trained an 8-billion-parameter masked diffusion model from scratch that rivaled a comparably sized autoregressive baseline and even handled some reversal-style tasks that trip up left-to-right models. Google DeepMind unveiled Gemini Diffusion at Google I/O in May 2025 as an experimental research model positioned as much faster than its fastest transformer at similar coding quality, and later released the open DiffusionGemma. On the commercial side, Inception Labs — founded by researchers behind foundational diffusion work — shipped Mercury, a production diffusion LLM whose later release was independently measured above a thousand output tokens per second. The 2026 consensus is coexistence rather than replacement: diffusion LLMs have earned a place for latency-sensitive and structured work, while the strongest general chat and reasoning models remain autoregressive.
| Platform | Behavior |
|---|---|
| LLaDA | A masking-based (discrete) diffusion model trained from scratch at the 8B scale, reported in early 2025. It was the clearest early proof that a diffusion model could match an autoregressive baseline of similar size in in-context learning and instruction following, and it challenged the assumption that core LLM abilities require left-to-right generation. |
| Gemini Diffusion (Google DeepMind) | An experimental research model unveiled at Google I/O in May 2025, generating blocks of tokens in parallel. Google positioned it as dramatically faster than its fastest autoregressive model at comparable coding quality. Access has been gated behind requests rather than shipped as a general product. |
| DiffusionGemma (Google) | An open diffusion model in the Gemma family that generates and refines blocks of tokens in parallel, giving developers open weights to run and study the approach rather than a hosted-only demo. |
| Mercury (Inception Labs) | A commercial, production-oriented diffusion LLM exposed via API and playground, aimed at low-latency, high-throughput use. Later versions were independently measured above a thousand output tokens per second at low cost per token — the strongest public evidence of diffusion's speed advantage. |
| Autoregressive models (for contrast) | GPT, Claude, Llama, and Gemini in its standard form are not diffusion models — they generate one token at a time. As of 2026 they still hold the top of the general reasoning and long-form quality frontier, which is why diffusion is a complement, not yet a replacement. |
The exciting thing about diffusion language models is not actually the speed number everyone quotes — it is that they break the assumption that good text has to be written left to right. That opens up genuinely useful behavior: infilling, planning globally, satisfying a fixed format, revising earlier words in light of later ones. Those are real advantages, and for latency-sensitive products the throughput is a legitimate leap. But the honest 2026 read is that this is a strong complement to autoregressive models, not their obituary, and anyone selling it as an LLM-killer is ahead of the evidence.
For anyone making content rather than building infrastructure, the practical takeaway is smaller than the hype suggests: which architecture drafts your text a little faster was rarely the thing standing between you and a full calendar. The bottleneck is turning any draft into finished, on-brand posts in every format and getting them scheduled. That is deliberately the layer a tool like Kompozy owns, and it is model-agnostic on purpose — it runs on Claude and OpenAI today and could adopt a fast diffusion model for drafting the day one is the better fit, because what reaches the person publishing is the content operation, not the decode speed. Watch diffusion LLMs closely; just do not confuse a faster token with a shipped post.
A diffusion language model generates text by iterative denoising instead of predicting one token at a time. It starts from a fully corrupted sequence and refines the whole thing in parallel over a series of steps until coherent text emerges — the same denoising idea behind AI image generators, applied to language. This lets it decode many tokens at once and use context in both directions, rather than strictly left to right.
GPT and Claude are autoregressive: one token at a time, each conditioned on the ones before it, so speed is bounded by sequence length. A diffusion LLM refines a block or sequence in parallel across denoising steps, so it can produce many tokens per pass with bidirectional context. The tradeoff is that diffusion models historically train less efficiently and can weaken coherence between tokens generated in the same step.
They can be, because parallel decoding removes the strict one-token-at-a-time bottleneck. Inception Labs' Mercury has been independently measured above a thousand output tokens per second, roughly an order of magnitude faster than comparably sized autoregressive models. Actual speed depends on hardware, task, and the number of denoising steps you run, so published figures are best-case rather than guaranteed.
Discrete diffusion corrupts text by masking or swapping tokens and learns to recover them — this fits language's categorical nature and is what most usable 2025–2026 models (LLaDA, DiffusionGemma, Gemini Diffusion) use. Continuous diffusion maps tokens to embedding vectors, adds Gaussian noise like an image model, then decodes back to tokens; it is harder to train but distills better into very-few-step samplers.
Not wholesale as of 2026. The strongest general chat and reasoning models are still autoregressive, and diffusion LLMs trade some peak quality and training efficiency for speed and bidirectional context. Their clearest wins are latency-sensitive and structured tasks. The realistic near-term picture is coexistence — including hybrid block-diffusion designs that borrow from both — rather than a clean replacement.