README.md

# Vext Labs

American AI research lab. We build **JUWEL**, a specialized AI that forms around you and stays on the record, and we open-weight the model lineage behind it.

## Open weights: JUWEL & GEMS

Our foundation bases are **JUWELS**. Our domain specialists are **GEMS**. Everything here is released under **Apache-2.0**, in **BF16** (no quantization), built on open **Qwen3-VL** foundations (Alibaba Cloud, Apache-2.0) and transformed with our own pipeline: refusal-direction abliteration plus organic layer-growth (CIP). We credit the foundation openly. The growth, the specialist fleet, and the tooling are ours.

### JUWELS (foundation bases)
- **JUWEL Garnet** — the origin stone; a retired flagship base
- **JUWEL Emerald** — the efficient general base
- **JUWEL Sapphire** — the vision-native base (144 layers)

### GEMS (specialist adapters)
Ruby (code) · Spinel (reasoning) · Aquamarine (math) · Peridot (science) · Topaz (engineering) · Lapis (language) · Pearl (education) · Jade (humanities) · Citrine (business) · Opal (creative) · Tourmaline (inquisitor) · Alexandrite (long context) · Sardonyx (legal) · Bloodstone (medical) · Tanzanite (finance) · Turquoise (reconciler)

Load a JUWEL base, then compose a GEM on top.

## Honest by default

- Weights are provided **AS IS** under Apache-2.0. Every model card states intended use, out-of-scope use, and base attribution.
- A **benchmark** appears on a card only once we can reproduce it on our own stack with a full audit trail. Where a number is not yet measured, the card says so plainly.
- We do **not** claim AGI, "never forgets," frontier parity, or from-scratch training. These are Apache-2.0 derivatives of Qwen3-VL, heavily modified.

## Also open (on npm)

`@vextlabs/theron-agent-sdk` (MIT) · `@vextlabs/theron-cli` (Apache-2.0) · `@vextlabs/stoa-verifier` (Apache-2.0)

## Links

Lab: [vextlabs.ai](https://vextlabs.ai) · Product: [juwel.ai](https://juwel.ai) · GitHub: [Vext-Labs-Inc](https://github.com/Vext-Labs-Inc)