Intelligence, measured.
An AI lab building its own line of models — starting with a hosted preview you can talk to today.
Product · Chat
Talk to models that are ours.
Chat is the browser workspace for the lab’s own models. It works the way the major chat apps do — pick a model, start a conversation, stream the reply — and it is plain about what exists: one hosted preview today, more models in training.
Streamed through a gateway
Replies stream through Alelyon’s authenticated gateway. The browser receives text only — never model weights, storage locators or provider credentials.
Plain about what exists
The model picker lists the public catalog. A preview says it is a preview, and models still in training cannot be selected.
Your conversation, your controls
Stop keeps what arrived; edit, regenerate and retry are yours to press. Nothing retries on its own, and the page saves no transcript.
- Hosted preview
- Closed weights
- Streaming replies
- Optional local runtime
Mission
Build models, and be plain about what they establish.
An answer is only as useful as its limits are clear. Alelyon keeps negative results beside the wins, writes an unmeasured state as unmeasured, and treats a refusal as a result.
To do that, we are building the whole stack ourselves:
- Interfaces
- A chat workspace in the browser, and Lattice for Windows — a workspace for choosing models, inspecting their declared structure and coordinating agents.
- Models
- Our own line of closed-weight models. One hosted preview is available today; the rest are in training.
- Compute
- The Alelyon Compute Kit runs expert training kernels — packed matrix products and clipped AdamW updates — on a selected Vulkan device.
- Evidence
- A public research record in which every note states its limit, and signed receipts a recipient can replay with their own inputs.
A receipt from that stack can be replayed against your own inputs with a public key obtained out of band. It detects revision of committed inputs, not whether the producer invented them.
How the lab worksLattice
A workspace for models and agents.
Lattice for Windows keeps the model choice explicit, reads a model’s declared structure without running it, and maps the agent sessions working at once. The build downloads with no sign-in. A hosted DQC-OS key is issued automatically when a signed-in account verifies with LinkedIn. An anonymous verification issues no key.
Actual product captures, from the project archive. Capture dates were not recorded. Scroll the image to inspect it on smaller screens.
Recent research
All notes- Separating model-authored programs from computed resultsThe restricted DSL parses an admitted program and computes its result through an interpreter instead of accepting a recalled figure.
- Evidence required before applying a receipt in another domainA new domain application needs a defined user decision, a covered uncertainty term, and evidence that the receipt helps with that decision.
- The scope of a storage-error execution boundThe execution certificate evaluates supported computations under a declared storage-quantization model and refuses unsupported cases.
Availability is not correctness. A caught exception, a non-empty dataframe, a bound port, or an HTTP 200 is not evidence that a result is right.