The fleet

Two tracks. Every model, honest status.

A general line (AGPT) and a dedicated agentic line (SIA), built on one owned, from-scratch spine. Here's what's built and serving today, then the full general catalog and the agentic line.

AGPT-Nano
general · from-scratch
Live

Our 110M from-scratch decoder — the live chat brain behind StretchGPT today.

Core · text-llm
AGPT-1 · Embed
stretch-embed
Live

~23M from-scratch encoder. Trained and serving; the foundation for semantic search.

Utility · text-embedding
AGPT · Code
mini · pro · max
Trained

Three sized code specialties — trained and registered in the catalog.

Code · text-llm
SIA-1.0
agentic generalist
Trained

Trained and measured — 95% overall, perfect tool-selection on our internal harness.

Agentic · tool-calling
Track 1 — General

The AGPT catalog.

One AGPT architecture, sized into Mini, Pro, and Max and specialized by domain — 28 models across eight capability lanes, graduating one at a time under governed sign-off. Status is updated from the live registry.

28Catalog models
4Trained & serving
8Capability lanes
2Governed / reserved

Statuses: Live — serving now · Trained — weights trained · Planned — on the roadmap · Governed — reserved, no build without sign-off.

Track 2 — Agentic

The SIA line.

A generational agentic line — one generalist, numbered by capability era, scoped to reliable tool-calling. The public name stays constant as the brain underneath grows from teacher-assisted to fully owned.

GenerationWhatStatus
SIA-1.0Agentic generalist — trained & measured (95% overall, 100% tool-selection)Trained
SIA-1.5Reinforcement-polished, sharper abstention, deeper harnessPlanned
SIA-2.0First fully-owned brain — our weights end to endPlanned
SIA-3.0+Scaled, self-improvingPlanned
See the architecture →

Want to use these models?

StretchGPT is the app where these models ship to work.

Open StretchGPT →