1 token a pass
16 passes“ years, to meet tax rules.↵↵Q: We place the sental sitted on the cand”
The careful fill: one token a pass, surest first.
Module 32 · Diffusion LMs
How does a model write in any order?
A diffusion language model starts from a reply of blanks and fills the surest first, several a pass. Here you run Larkspur’s toy masked model in a type frame, beside Larkspur mini’s typewriter, and see what fewer passes cost.
Read as
The whole idea in one example
One phrase, filled in 16 passes and in 1
16 tokens a pass
1 pass“ years, to meet tax rules.↵↵Q: Wo the the a a a a a a the the thesee”
1 of 16 tokens the same as the careful fill; 9 repeats.
Both run Larkspur’s toy masked model, Larkspur mini’s size, trained once on Larkspur’s documents. It shows the mechanism; its replies are not what a production model writes.
How it works
Six ideas carry the module
1 One word at a time
A left-to-right model takes a pass per token
Larkspur mini continues “ safety boots and high-visibility” with “ reports.↵↵↵Q: Can I order? A: Yes”, one token a pass.
2 Fill the blanks
A masked model guesses every blank at once, reading both sides
MDLM (2024) found simple masked diffusion works better than previously thought, approaching left-to-right models’ perplexity, and can be sampled block by block for text of any length.
Source, checked 25 October 2026
3 Sure words first
Each pass keeps its surest guesses
The LLaDA paper (2025), on an 8-billion-parameter masked diffusion language model, reports that low-confidence remasking (keeping the surest guesses) consistently outperforms random remasking.
Source, checked 25 October 2026
4 Faster and worse
More tokens a pass: fewer passes, worse words
Fast-dLLM (2025) decodes only the tokens above a confidence threshold in parallel and adds an approximate key-value cache, reporting up to 27.6× more throughput with minimal accuracy loss.
Source, checked 25 October 2026
5 The length is fixed
The frame’s size is set before the first pass
LLaDA sets the reply’s length before sampling starts: it begins from a fully masked sequence of that length, and drops whatever comes after the end-of-text token.
Source, checked 25 October 2026
6 Reading a speed claim
Tokens a second = tokens a pass × passes a second
Inception, the vendor, claims its diffusion model Mercury 2 writes 1,009 tokens a second on NVIDIA Blackwell GPUs. The vendor’s own figure, from its blog.
Source, checked 25 October 2026
Everything here runs live in your browser: Larkspur mini, the left-to-right baseline, and the toy masked model, which fills the frame a pass at a time.
The story
A vendor’s speed, read by Larkspur
A vendor promises a model that writes 1,009 tokens a second by filling a reply’s blanks in any order. Larkspur Supply Co. wants to know what that buys. You’re on the team that decides how many tokens a pass each kind of reply can take.
Diffusion models showed pictures coming out of noise, and Transformers showed a model writing left to right. Here text comes out of blanks.
- 10
- pinned phrases every setting is tried on
- 16
- passes for the careful fill of a 16-token reply
- 1,009
- tokens a second: the vendor’s own claim
- 22
- passes, the fewest that clear every bar on the workbench; the vendor’s default takes 6 and fails
Your path
The chapters
Chapter 6 · done
Reading a speed claim
What makes up tokens a second, and what should you ask about it?
Chapter 7 · done
Workbench: speed per kind of reply
How many tokens a pass, and which remasking, does each kind of reply need?
Who it’s for
- Engineers
- Product managers
- Analysts
- No code needed
Product managers, engineers and analysts weighing a faster model. Diffusion models and Transformers come first; no code or maths needed.