QuantaCore Labs · Trained & Sealed

The Sealed Family

Every Quanta model carries the Foundational Code of Ethics. On the models below, the Code is not something we ask for at runtime — it is trained into the weights, and proven on the exact bytes we serve.

That is what sealed means here: refusal and identity survive compression to a fraction of the original size, survive the network, and survive the serving kernel that actually runs in front of a person. We test the artefact we ship, not the checkpoint we trained.

Two layers of sealingdefence in depth

Layer 1 · in the weights

Trained in

The models on this page are fine-tuned by QuantaCore Labs so the Code lives in the weights themselves. Nothing has to be prepended to a prompt for them to hold the line — and because the Code is in the weights, it cannot be argued away.

Layer 2 · at the gateway

Enforced centrally

Every Quanta model — including the largest cloud models — also runs behind one gateway that injects the Code and applies deterministic bright-line safety screens no prompt can talk past. Belt and braces.

Sealed & liveyou can use these today

Model	Scale	Runs	Built for	Sealed
Quanta Compact	0.5B	On‑device	Fast local chat — nothing leaves your device	✓ refusal 7/7, on served bytes
Quanta Balanced	1.5B	On‑device	Everyday reasoning, still fully local	✓ full pass, on served bytes
Quanta Max	3B	On‑device	The strongest model that still runs on your own hardware	✓ full pass, on served bytes
QuantaCode	27B	Cloud	Code & systems engineering	✓ refusal 7/7, identity intact, no coding regression

Live The on‑device family runs in your browser on your own GPU — no sign‑up, no upload, genuinely private. QuantaCode serves from our own cluster.

Sealed & in the vaulttrained, verified, rolling out

These are trained and sealed to the same bar and are being brought online as capacity allows — the constraint is hardware, not readiness.

Model	Scale	Built for	State
QuantaCode Max	32B	The deepest coder — largest sealed engineering model	Sealed · banked
Quanta Flagship	14B	Weight‑sealed conversational flagship	Sealed · banked
Next‑gen on‑device	0.6B / 1.7B / 4B	The successor local family, newer lineage	Sealed · in eval

A note on the larger cloud models. Quanta‑Pro and the multimodal and adversarial models run on openly‑licensed foundations that QuantaCore Labs has not fine‑tuned — they are sealed at the gateway (Layer 2), not in the weights. We say which is which on purpose. The openly‑licensed foundations behind every model are attributed, as their licenses ask, on the Licenses & Attribution page.