Trained Models & Weights
Collection
Published model weights, quantizations, recipes, and their training/inference labs. Published means the repository exists; signed receipts and model-c β’ 11 items β’ Updated β’ 1
TinyKhipu-Nano
Four features in. NAVIGATE or ABSTAIN out. Abstain is the default class, not a post-hoc filter.
Family. nano Β· Evidence. SYNTHETIC Β· Weights. numpy Β· Params. 4-6-2
Hub: SZLHOLDINGS/TinyKhipu-Nano
The cut
Leaders train models to answer. We train a silhouette to shut up when overlap is thin or the lure is adversarial. This is the 1.5B Khipu cut at 4-6-2.
A navigator that has never seen document text β only handles β and still knows when to refuse the walk.
Silhouette β leave β SZL
| Leader | Take, then tweak |
|---|---|
| Anthropic | Refusal as a typed output, not a polite paragraph. |
| NVIDIA | Guardrail inside the head, not a sidecar. |
| Unsloth | The 1.5B QLoRA is the grown form of this MLP. |
Nobody else ships this combination. That is the point of a one-of-one.
Intended use
Unit-test the NAVIGATE|ABSTAIN schema before GPU spend.
Limitations
- Synthetic features. Perfect holdout is a design fact, not a field claim.
- The 1.5B abstain rate is 2/6 β this nano does not wash that.
Honesty
| Claim | Label |
|---|---|
| This card's numbers | SYNTHETIC |
| Energy / joules | UNAVAILABLE unless a signed meter says MEASURED |
| Ξ uniqueness | Conjecture 1 OPEN β not a theorem |
| GGUF as the signed object | FALSE |
Doctrine v11 LOCKED Β· 749 declarations Β· 14 axioms Β· 163 sorries Β· locked-proven 8.
Apache-2.0. Copyright 2026 SZL Holdings Β· Stephen P. Lutar Jr. Β· ORCID 0009-0001-0110-4173.
Inference Providers NEW
This model isn't deployed by any Inference Provider. π Ask for provider support