Models

AiQL runs a family of pre-trained models, each specialized for one job. Convert uses layout, OCR, and speech-to-text. Shape and extract use pipeline language models. Queries use answer models over embeddings of the knowledge layer. Observatory also uses text-to-speech and speech-to-speech.

Parameters

These are the scales in production.

Job	Kind	Scale	Window
Layout	Object detector	33 million parameters	One rasterized page
OCR	Compact vision-language model	0.9 billion parameters	One region crop
Pipeline	Sparse mixture-of-experts	35 billion total, 3 billion active per token	262k tokens
Answers	Sparse mixture-of-experts, vision-language	35 billion total, 3 billion active per token	262k tokens
Embeddings	Dense text encoder	4 billion parameters	32k tokens, 2560-dimension vectors
Speech-to-text	Language-model ASR	Multilingual	Audio and video into convert
Text-to-speech	Neural speech synthesis	Low-latency	Up to 2k input tokens, `mp3`
Speech-to-speech	Realtime duplex audio-language	Speech, text, and images in; speech and text out	32k tokens

Layout

Finds regions on each page before any text is read.

OCR

Reads those regions, with dedicated passes for tables, formulas, charts, and seals.

Pipeline

Infers or applies a schema, then extracts entities and relationships.

Answers

Reasons over the workspace knowledge layer in ask, reason, or build mode.

Embeddings

Vectors for queries, schema names, and instances. Retrieval fuses exact, lexical, and semantic match.

Speech-to-text

Transcribes audio and video into convert.

Text-to-speech

Speaks Explore replies in Observatory.

Speech-to-speech

Realtime duplex voice over workspace knowledge.

Applications

Platform

SDKs

Models

Parameters

Layout

OCR

Pipeline

Answers

Embeddings

Speech-to-text

Text-to-speech

Speech-to-speech

​Parameters

Layout

OCR

Pipeline

Answers

Embeddings

Speech-to-text

Text-to-speech

Speech-to-speech

Parameters