A classic language model writes like someone typing: one word, then the next, each depending on what came before. A diffusion model writes like a sculptor: it starts from a shapeless draft of the desired length, and refines it progressively until a text emerges. The second approach produces everything in parallel, and that is what changes its speed.
We have explained diffusion models for images. Their application to text is more recent and is starting to produce interesting results.
Why classic models are slow
A usual language model is autoregressive: it produces a token, adds it to what it has already written, and starts again for the next one.
This dependency is what makes its quality — each word takes account of all the previous ones — and its limit. The process is sequential: it is impossible to compute the tenth token before the ninth.
This is why serving a model requires particular hardware, as we explained regarding chips dedicated to inference. The limiting factor is not computing power but the speed at which one can chain together steps that wait on one another.
How diffusion proceeds
The approach turns the problem on its head. The model starts from a sequence of the desired length, filled with noise, and refines it over several passes.
At each pass, it looks at the whole and improves all positions simultaneously. After ten or twenty passes, a coherent text has emerged.
The gain is structural: each pass processes everything in parallel, which hardware does very well. Where a classic model chains a thousand steps for a thousand tokens, a diffusion model does twenty for the same length.
This explains the announced throughputs, in the order of a thousand tokens per second on ordinary hardware, well above what comparable autoregressive models achieve.
Speed is not a comfort. It decides what is possible. We described in our article on voice that dropping below 300 milliseconds turns an exchange into a conversation. For an agent chaining dozens of steps, each faster step compounds. And for code completion, where a suggestion is expected while you type, latency is the main criterion — well ahead of absolute quality.
The limits, which explain why it is not widespread
Length is fixed in advance. The model works on a sequence of a given size. Producing a response whose length is unknown is less natural than with a model that stops when it has finished.
Coherence over long texts. The autoregressive approach guarantees that a word takes account of everything before it. In diffusion, this coherence must be obtained through successive passes, and it degrades faster on long texts.
Step-by-step reasoning. This is the most important limit. What we described in our article on reasoning models rests precisely on chaining: the model writes a step, reads it, deduces the next. A mechanism that produces everything simultaneously sits badly with that logic.
Where things stand today
Recent text diffusion models claim a level comparable to the fast models of the major families — not to the leading-edge models.
This is consistent with their positioning. They do not aim to replace a reasoning model on a difficult problem, but to occupy the segment where latency comes first: completion, text transformation, classification at volume, real-time assistance.
This is exactly the routing logic we recommend in our guide to model sizes: the question is not which is best, but which is enough for that particular task.
What to take away
This approach is interesting less for what it is than for what it signals: after years in which the whole industry built the same architecture, alternative paths are reappearing.
This is consistent with what we described in our article on scaling laws: when you can no longer win by going bigger, you look elsewhere. Mixture of experts, inference-time reasoning, distillation and now text diffusion all belong to the same period — the one in which ingenuity replaces force.