Status
Live
Most AI is rented.
Ours is owned.
StretchAI is the artificial-intelligence practice of StretchGroup Technologies Inc. We build our own models across two tracks — general intelligence and a dedicated agentic line — and we run them ourselves. StretchGPT's live chat already answers on a model we built and tuned ourselves, served on our own hardware.
Two tracks
One owned spine. Two jobs.
We build in two model lines on the same from-scratch foundation — general intelligence, and a dedicated agentic line that uses tools.
Track 1 — AGPT · General
chat · reason · specialize
Our general line. AGPT-1 turns text into vectors; AGPT-Nano is our 110M from-scratch decoder; and a full fleet of specialists covers code, business, vision, voice, and imaging.
Routing layer · skill modules · own runtimes
Track 2 — SIA · Agentic
tools · bounded agency
A generation of agents that call tools reliably. SIA-1.0 is trained and measured (95% overall on our in-development harness), scoped to tool-calling, with our own weights growing in every generation.
Canonical tokenizer · verified tool data · SFT → RL
Flagship · Live
There's no one else's model underneath it.
AGPT-1 is a text-embedding model we built from nothing. We wrote the tokenizer, designed the architecture, and trained every weight ourselves, on data we control. It's trained and serving today — the embedding model we built to bring semantic search to the platform — and the intellectual property is entirely ours, with no base-weight license to answer to.
Learn more →Type
Text embedding384-dim
Architecture
Encoder~23M params, built in-house
Tokenizer
Byte-level BPE32k vocab, our own
Training
MLMFrom random init
Ownership
100% ownedNo base-weight license
The fleet
Every model we've built. One place.
Seventy-seven models and iterations across our lines — from-scratch AGPT, agentic SIA, foundation SFM, and more. Some are live or trained today; for the rest, including the retired ones, we publish exactly where each one stands. Here's a sample.
Cognitive suite
Every answer is checked before it ships.
A layer of deterministic checks runs around each model response: it holds context, critiques the draft, verifies claims against sources, and enforces safety. The checks are governed and auditable, and they fail open — they sharpen the answer without ever standing between a customer and a reply.
25Cognitive algorithms
23Wired into live chat
100%Deterministic & governed
Fail-openNever blocks a response
Roadmap
Nothing ships before it's ready.
Every model follows the same path from candidate to production, and we don't move the goalposts. Today the Code family is trained and AGPT-1 is live. The general-purpose core comes next.
Built · weights
24
Planned
45
Governed / retired
8
Total
77