Introducing Phron · a Super-Diffusion language model

Frontier intelligence, trained sovereign.

Phron does not write an answer one word after another. It draws the whole of it out of noise at once, and sharpens every line together, pass by pass, as an engraver works a plate.

See the prototype How it differs

Plate I.The Catacombs. A gallery of the sequential age, wherein each word waited upon the last.

From the sequential to the parallel

Every language model of the present day sets its words as a scribe does: one, and then the next.

A sequential model cannot write its tenth word until it has committed to the ninth. Every word is a wait, and long answers grow long in the waiting. The whole trade has grown so accustomed to this that it builds its clocks around it, and calls the delay natural.

Phron is built on Super-Diffusion. It begins with the entire reply as an undifferentiated field, a page of noise, and in each pass sharpens all of it together. Early words are settled in the light of later ones. A clumsy phrase in the first line may be reworked while the last line is still forming. After a small number of passes, the whole answer reads true at once.

The consequences are practical. The reply arrives in a handful of passes rather than a procession of words. The cost of an answer no longer grows word by word. And the model may reconsider what it has already written, which a sequential decoder, by its nature, cannot.

Parallel: Every position in the reply is refined in the same pass, not in turn.
Coarse to fine: The sense settles first, the particulars after, as a plate is worked from broad strokes to the burin.
Revisable: Any word may be reconsidered until the page is fixed, so the beginning agrees with the end.
Sovereign: Trained and served under your own roof, understandable and dependable to those who run it.

Plate II.Fig. 1, the sequential press, which sets one sort after another. Fig. 2, the Super-Diffusion plate, which begins as noise across the whole page and settles it in a few passes.

The Diffusion Press · a prototype

Set a question in type, and watch the two presses run on the same passage.

Below, the sequential press and the Super-Diffusion plate are put to the same passage at the same instant. The one sets a word and then the next. The other begins with the page as noise and sharpens every position on each pass, so the sense appears everywhere at once and is then made particular.

The Diffusion Press

Choose a passage, then set it in type. Both presses start together.

The Sequential Press

0.0 s

awaiting the first sort

The Super-Diffusion Plate

0.0 s

the plate is blank noise

The passages are set in advance, and the clocks illustrate the difference in kind between the two manners of decoding. Nothing above is a measurement.

Phron

Built for depth over spectacle.

Phron is strongest where context runs long and the stakes are technical: writing and repairing code, sustaining multi-step agentic work, and reasoning carefully in specialist domains. Super-Diffusion is what lets it hold a long answer whole, rather than committing to each word before the shape of the reply is known.

It serves through an OpenAI-compatible inference API and is built for both research and production deployment. Developer access is opening in stages — write to us for early availability.

Code: Generation, editing, and repair in real codebases
Agents: Multi-step reasoning across long tool-use trajectories
Domains: Code, agents, finance, medical, and science
API: OpenAI-compatible inference for research and production

Research

What makes a diffusion language model genuinely useful for expert work?

We study how a model sustains reasoning over long contexts, how agentic behaviour holds together across many steps, how a reply may be revised while it is still forming, and how deep domain competence emerges. What we learn is shipped into Phron.

Code: Generation, editing, and repair across large repositories
Agents: Long-horizon reasoning over extended tool-use trajectories
Finance: Quantitative and regulatory language understanding
Medical: Clinical and biomedical reasoning under precision constraints
Science: Technical literature comprehension and hypothesis support

Plate III.The Ossuary of Reason. The vault beneath the laboratory, where the sequential age is laid to rest, by candle and by the daylight of the stair.