Why this matters now

Watch any AI chatbot answer a question and the text arrives a word or two at a time, as if someone were typing. That isn't a stylistic choice. It's literally how ChatGPT, Claude and Gemini write: each word is chosen only after the one before it.

There's another way. AI image generators don't paint a picture pixel by pixel; they start with the whole canvas as static and sharpen all of it at once. Diffusion language models (dLLMs) apply the same idea to text. They start with a blank answer, fill in the whole thing in rough form, and refine it over a handful of passes.

For years, the editor was the fast but weaker option: much quicker, noticeably less capable. That changed on October 19, 2026, when the startup Inception released Mercury 3. It says the model writes about 1,068 tokens (roughly, words) a second and beats fast models from OpenAI and Anthropic on 8 of 12 tests. If those numbers hold up, the way AI writes could start to change, beginning wherever speed matters most: voice assistants, coding tools and AI agents.

What's in this post

• Part Two explains how today's chatbots write, and why that approach is hitting a wall.
• Part Three explains diffusion: how a model can fill in a whole answer at once, and what it gives up to do so.
• Part Four looks at Mercury 3 and the other companies building these models.
• Part Five looks at where this is heading.

Writing one word at a time

ChatGPT, Claude, Gemini and nearly every chatbot you've used work the same basic way. They write an answer one word at a time, from start to finish. Technically they work in tokens, which are small chunks of words, but "words" is close enough.

At every step, the model asks one question: "Given everything written so far, what's the most likely next word?" It ranks its best guesses, picks one, adds it to the answer, and asks the question again. It repeats this hundreds or thousands of times until the answer is complete.

This approach is called autoregressive, a fancy way of saying "each new piece depends on the pieces before it." Press play to watch it write a sentence.

It looks almost too simple. Here's why it has worked so well.

Why it works so well

• Each step is a simple question. Predicting just the next word is a manageable task, and it gives the model a clear, well-defined job.
• It's easy to train. Every sentence on the internet is a ready-made practice example: hide the next word and see if the model can guess it.
• It matches how we read and reason. Step-by-step thinking, where each step builds on the last, fits naturally into a one-word-after-another process.

This approach has powered every major AI chatbot since ChatGPT launched in late 2022.

That same one-word-at-a-time design is also what holds it back.

The bottleneck it creates

Problem 1: it's slow, and it can't be done in parallel

Word 50 can't be written until word 49 is done. A 1,000-word answer takes 1,000 separate steps, one after another. No matter how powerful the computer, it can't skip ahead.

Problem 2: the computer sits mostly idle

This is the less obvious problem. To produce each single word, the computer has to load the entire AI model, billions of numbers, from memory, do a tiny amount of maths, and then do it all over again for the next word.

The real limit is how fast data can be moved around, not how fast the chips can calculate. That speed has a name: memory bandwidth, the amount of data that can travel between a chip's memory and its calculating cores each second. In the analogy, it's the width of the road between the warehouse and the kitchen. A faster chef doesn't help if the road stays the same; a wider road does.

Why the whole industry is fighting over memory

Chipmakers widened the road by moving the warehouse next door. High-bandwidth memory (HBM) stacks memory chips on top of each other like a tower and places them right beside the processor, connected by thousands of tiny wires instead of a few long ones. NVIDIA's newest AI chip, Rubin, carries eight of these towers and can move about 22 terabytes a second, roughly the contents of a few thousand feature films.

Because every chatbot answer is limited by that road, AI companies want as much HBM as they can get, and only three companies make it: SK hynix, Samsung and Micron. All three say their supply is largely spoken for, and Micron has told investors its HBM for next year is effectively sold out. Prices are expected to more than double in 2027, and the squeeze has spread to ordinary memory, pushing up PC prices. Earlier this month, AMD's CEO Lisa Su flew to Seoul to press Samsung for more supply.

This is where diffusion comes in. A typist makes one trip down the road per word. An editor brings back many words per trip, so it gets far more out of the same, scarce, expensive memory.

Problem 3: no take-backs

Once a word is written, it stays. If the model starts a sentence badly, it has to keep building on that mistake. It also has to commit to how an answer begins before it "knows" how it will end.

Writing like an editor

The idea comes from image generators

If you've used an AI image tool like Midjourney or Stable Diffusion, you may have seen the picture appear as a blurry mess that sharpens over a few seconds. That's diffusion. The model is trained by taking real pictures, adding visual "static", and learning to remove it. To create a new image, it starts with pure static and cleans it up step by step. Crucially, it works on the whole image at once, not pixel by pixel.

Applying it to text

Text is trickier, because you can't "blur" a word. Researchers solved this by treating blanked-out words as the equivalent of static.

1. Training: take real sentences, blank out a random portion of the words, and teach the model to fill in the blanks. Sometimes a few words are hidden, sometimes almost all of them.
2. Writing an answer: start with a completely blank answer of a set length. Then the model guesses every blank at the same time, keeps the guesses it's most confident about, and erases and retries the uncertain ones. It repeats this for a handful of passes until no blanks remain.

Here's the same sentence the typist wrote, written by an editor:

That raises the most natural question about diffusion: if the model starts with nothing but blanks, how can it be confident about any word?

How it knows what to write, and which words to trust

The answer has three parts.

1. The blanks are never the real starting point. Your request is.

A diffusion model never writes in a vacuum, and neither does a traditional chatbot. When you type a request, it sits in front of the blanks the whole time, fully visible, on every pass. The sentence in the animation would only appear because someone asked for it, for example:

The intent comes entirely from that request. The model didn't decide to write about a robot; the request told it to. With a different request, the same blanks would fill with completely different words.

2. "Confidence" just means "one answer stands out"

For every blank, on every pass, the model produces a ranked list of possible words with a likelihood for each, just like the typist does for its next word. A word counts as confident when it's far more likely than any alternative. Look at the robot request again:

• "robot" is practically dictated by the request. Almost no other word fits, so it's confident right away.
• "garden" and "watered" are strongly implied ("takes care of a garden"), so they're confident too.
• The full stop at the end is an easy call for a one-sentence request.
• "little", "quietly" and "before sunrise" are much less certain. "Early morning" could become "at dawn", "before sunrise" or "early one day", and a children's-story robot could be little, small, friendly or old. No single option stands out.

The model keeps the clear winners and leaves the uncertain blanks open. That's why the editor in Figure 2 locks in "robot", "watered", "garden" and the full stop first: they're the words the request practically spelled out.

3. Every word it keeps makes the remaining blanks easier

Once "robot", "watered" and "garden" are in place, the gaps between them get much easier to fill. "The ___ robot ___ watered ___ ___ in the garden" leaves very few sensible options for the connecting words.

This is also how the model "changes its mind". On an early pass, "flower" might look about as good as "plant". Once more of the sentence is settled ("watered every ___"), "plant" becomes the better fit, so the rough guess gets replaced before it's locked in.

Where does this judgement come from?

From training. The model has practised filling in blanks on an enormous amount of real writing, with blanks in every position and at every level of emptiness. Over billions of examples, it learned which words tend to appear given a certain request and a partly finished sentence. It doesn't "understand" intent the way a person does; it has learned, very precisely, what usually fills which gaps.

The editor can change its mind about a word. But can it move words around, the way a person editing a draft would?

Can words move around?

In most diffusion models today, positions are fixed. Think of the blank answer as a row of numbered boxes, one word per box. When the model decides "robot" goes in box 3, "robot" is tied to box 3. The model fills boxes; it doesn't slide words between them, swap their order, or squeeze a new box in between two others.

• The position is always locked. A word kept in box 3 stays in box 3.
• The word itself is usually locked too. In the most common design, once a word is kept, it's final; only the uncertain boxes get rethought. Some newer designs let the model reopen a kept box and change its word if later context shows it was a mistake. Even then, it changes what's in the box, not where the box is.

Why this is a real limitation

Imagine the model locks "garden" in box 10 early on, then realises the sentence would read better as "watered every plant in the vegetable garden". There's no free box between "the" and "garden". It can't push "garden" one box to the right, so it has to make the sentence work with the boxes it has.

What about answer length?

Because the number of boxes is set in advance, the model needs a way to write shorter answers. It uses a special "end of text" filler: if a request needs only 9 words but the model was given 13 boxes, it fills the last 4 with "nothing here". Longer answers are usually written in chunks: finish one block of boxes, then start the next.

How researchers are working around it

• Writing in chunks, sometimes called block diffusion. The model fills one short block at a time, left to right, so an early mistake only constrains a small block rather than the whole answer.
• Insert-and-delete models. Newer research designs let the model add or remove boxes as it refines the draft. This works more like real editing, but it's harder to build and not yet common in released products.

With those limits in mind, here's what writing like an editor buys, measured against the typist's problems from Part Two.

What it fixes

The biggest win is the number of trips through the model. Look back at the "Trips" bars in Figures 1 and 2: the typist needed 13 trips to write the 13-word sentence, one per word, while the editor needed 4 passes. That's the same sentence in about a third of the time.

Diffusion isn't free, though. It swaps the typist's problems for new ones.

The new trade-offs

• Guessing several words at once is harder. When the model fills two blanks at the same time, neither guess knows what the other will be. If the answer should be either "New York" or "Los Angeles", it might fill the two boxes independently and produce "New Angeles". Avoiding this takes more passes, which eats into the speed gain.
• Quality has historically trailed. Model for model, diffusion has usually scored a bit lower than the best one-word-at-a-time models, though the gap is shrinking fast (see Part Four).
• Answer length is awkward. The row of boxes from Can words move around? has to be sized in advance, and getting that size right is still unsolved in some designs.
• The speed advantage shrinks at large scale. A big provider serving thousands of users can bundle their requests together, which already keeps its computers busy. Diffusion's edge is biggest for one user who wants an answer right now, and smaller in a crowded data centre.
• Answers arrive in chunks, not a smooth stream. Chat apps that show text appearing word by word don't fit as naturally.
• Less mature tooling. Years of industry effort have gone into training and serving one-word-at-a-time models. Diffusion is still catching up.

What Inception released

Inception is a startup that builds only diffusion language models. It was founded by Stefano Ermon, a Stanford professor who helped develop the diffusion technique behind today's image generators. On October 19, 2026, it released Mercury 3, which it calls the most capable diffusion language model ever released.

Inception says Mercury 3 is more than 40 times larger than its first Mercury model, and that its benchmark scores roughly tripled over that span. The headline claim is that "text diffusion scales": making the model bigger makes it smarter, the same pattern that drove progress in traditional chatbots.

• Speed: about 1,068 tokens per second on NVIDIA chips.
• Memory: up to 1 million tokens of input at once, roughly several long books, up from 260,000 in Mercury 2.5.
• Price: $0.20 per million input tokens and $0.70 per million output tokens at list price, with a 50% launch discount ($0.10 and $0.35).
• Availability: Inception's own platform, plus OpenRouter, Baseten and Vercel. It works with apps already built for OpenAI's interface.
• Early uses: web-browsing agents, tutoring, real-time search and customer support.

Inception put Mercury 3 head to head with two of the leading "fast and affordable" traditional models.

How it compares

The comparison is against OpenAI's GPT-6 Luna and Anthropic's Claude Haiku 5.5, both one-word-at-a-time models.

Speed (Inception's own measurements)

That's roughly 4 times faster than Haiku 5.5 and 8 times faster than GPT-6 Luna.

Quality (selected benchmarks, Inception's own tests; higher is better)

Across the 12 tests Inception published, Mercury 3 came out on top in 8. Claude Haiku 5.5 led on longer coding tasks and long-document recall, and GPT-6 Luna led on two knowledge and long-context tests.

Price: at the launch discount, Mercury 3 matches both rivals on input cost and is cheaper on output. At full list price, its output cost ($0.70 per million tokens) is higher than both rivals ($0.50). Inception hasn't said when the discount ends.

If those numbers hold, they change the story about diffusion. That's a big "if".

Why it matters, and the caveats

For years, the story of diffusion text models was "much faster, but noticeably less smart". Mercury 3 is the first claim that a diffusion model can be both much faster and as good as or better than leading speed-focused models from the biggest AI labs. If it holds up, the next question is whether diffusion can reach the level of the very best models, not just the fast tier.

Inception is the purest diffusion bet, but it isn't the only company building these models.

Who else is building them

• Google DeepMind showed an early demo, Gemini Diffusion, in May 2025. In June 2026 it released DiffusionGemma, a free, downloadable model that runs on a high-end home computer and generates text up to about 4 times faster than its traditional sibling. Google is upfront that it scores lower on quality, and still recommends its regular Gemma 4 models when quality matters most.
• NVIDIA released Nemotron-Labs-Diffusion in May 2026, a family of free models in several sizes. One model can switch between both styles: in one mode, the diffusion side writes a fast rough draft and the one-word-at-a-time side double-checks it. NVIDIA also showed a way to convert an existing traditional model into a diffusion model, rather than building one from scratch.
• ByteDance (TikTok's parent company) and Ant Group have released diffusion models too: Seed Diffusion and the LLaDA series. Universities including Stanford and Cornell continue to drive much of the core research.
• OpenAI and Anthropic haven't publicly released diffusion language models. Their flagship products still write one word at a time.

Where the industry is heading

Blending, not replacing

The likeliest future isn't one approach beating the other. The two are already merging: writing in chunks one after another but filling each chunk all at once; a fast diffusion draft followed by a careful traditional check, as NVIDIA's Nemotron already does; and single models that switch between modes depending on the task.

Speed-hungry uses will adopt it first

• Voice assistants, where even a one-second delay feels awkward. Inception already has a voice-focused version, Mercury Voice.
• Coding tools, where faster drafts mean more chances to try, test and fix.
• AI agents that browse the web or complete tasks by making hundreds of quick AI calls behind the scenes. Small delays add up fast.
• AI on your own device, where diffusion makes better use of a single computer's chips.

Everyday chat and the most demanding reasoning tasks will likely stay with traditional models for a while longer.

A possible hidden advantage: learning more from less

AI companies are starting to run short of high-quality text to train on. Some research suggests diffusion models learn more from the same amount of reading, because each sentence can be blanked out in many different ways for practice. If that holds at scale, it could matter even more than speed.

Pressure on the big players

A model that's several times faster at similar quality and price puts pressure on OpenAI, Anthropic and Google's "fast and cheap" offerings. Expect them to respond with lower prices, faster versions of their own models, or by adopting diffusion-style techniques themselves.

Text, images and video converging

Text AI is borrowing diffusion's speed from image AI, while video AI is borrowing the step-by-step approach from text AI so it can make longer videos. Over time, the same mixed approach may power AI that creates words, images, video and sound together.

Bottom line

Diffusion language models write the way an editor revises a draft, rather than the way a typist types. They're much faster and handle editing and whole-answer planning well, at the cost of some new technical challenges. Until recently they lagged in quality. Mercury 3 claims to have closed that gap against leading fast models, but those claims still need independent confirmation. The most likely future is a blend of both approaches, with diffusion leading wherever speed matters most.

Glossary

Autoregressive

Writing one piece at a time, where each new piece depends on the ones before it. How today's chatbots write.

Benchmark

A standard test used to compare AI models, such as solving science questions or fixing software bugs.

Block diffusion

Writing in short chunks, one after another, but filling in each chunk all at once.

Diffusion

A way of generating content by starting from noise (or, for text, blanks) and refining the whole thing over several passes.

Diffusion language model (dLLM)

A text model that writes by filling in a blank answer over a few passes, rather than one word at a time.

High-bandwidth memory (HBM)

Memory chips stacked into towers and placed right next to an AI processor, so data can move between them much faster. Made by SK hynix, Samsung and Micron, and in short supply.

Memory bandwidth

How much data can move between a chip's memory and its calculating cores each second. For one-word-at-a-time models, it's usually the real speed limit.

Pass

One trip through a diffusion model, in which it guesses every remaining blank and keeps the most confident guesses.

Token

A small chunk of text, often a word or part of a word. Models read and write in tokens.

Tokens per second

How fast a model writes. Mercury 3 claims about 1,068; leading fast traditional models write roughly 100 to 250.