Request Q-Prime evaluation access
Q-Prime is a commercial quantum-structured embedding model for regulated AI.
Model weights, adapters, and training data are not distributed.
Approved evaluation requests receive a time-boxed API key for non-production
use under the QGI Commercial Model License v1.0 §3 (90-day evaluation grant).
Production use requires a Permitted Commercial License — contact@qgi.dev.
QGI reviews all requests manually. Evaluation access is time-boxed and
non-production. Target SLA: 3 business days.
Log in or Sign Up to review the conditions and access this model content.
Q-Prime — QGI Quantum-Structured Embedding Model
No weights on this repo. Q-Prime is a managed API. This Hugging Face page exists for discovery, procurement, and citation. There is no
.safetensors,.bin,.onnx,.gguf, or any other loadable artefact here — by design. Access is granted via API key after approval. Jump to How to access ↓
Q-Prime is a quantum-structured embedding model purpose-built for regulated AI. It powers the QAG engine — Quantum-Augmented Generation — QGI's successor category to classical RAG for applications that cannot afford to hallucinate.
Q-Prime is accessed as a managed API. Weights, adapters, tokenizer, and training data are not distributed. See How to access below.
License: QGI Commercial Model License v1.0 — evaluation access available on request via the Request access button above; paid commercial license required for production. See LICENSE.md for full terms. Licensing:
contact@qgi.dev.
What Q-Prime does
Rules and regulated text carry structure that classical embeddings discard. A clause holds several meanings at once; it is correlated with other clauses in ways that reinforce, condition, or contradict. A cosine-based retriever flattens all of that into a single point in a vector space and loses the signal that matters.
Q-Prime is built on the opposite premise. It finds entangled superpositions in rules and text and emits a quantum-structured representation that keeps them intact. The representation is more compact than a classical embedding — relational structure lives in the state itself, not padded into extra dimensions — and it exposes parameters that cosine similarity cannot see: polarity, scope, conditions, obligation, and cross-rule dependency.
Q-Prime feeds a pipeline — the QAG engine — that reads this structure at inference time. Interference between related representations produces a signed signal: same-polarity related statements reinforce, opposite-polarity related statements cancel. The sign is the decision. Contradictions that differ only by a negation — "must report" vs "must not report" — become separable, and the same mechanism surfaces scope conflicts, conditional overrides, and other parameters that classical retrieval averages away.
Real quantum formalism on classical hardware
Q-Prime uses the mathematical apparatus of quantum mechanics — Hilbert-space states, superposition, interference, the Born rule $P(\text{outcome},|,\psi) = |\langle\text{outcome},|,\psi\rangle|^2$ — evaluated on commodity GPUs. It is not a quantum-hardware model and it is not "quantum-inspired". The operator algebra is the real one, the Born rule is the real one, and the probabilities it emits are calibrated in the same sense that physical Born-rule measurements are.
This is what lets Q-Prime expose a Born-rule classifier $\arg\max_c |\langle c,|,\psi\rangle|^2$ as a zero-shot categorization primitive, and a signed interference signal $\mathrm{polarity}(a,b)\cdot\langle a,|,b\rangle$ as a conflict-detection primitive.
Intelligence signals (public API surface)
Q-Prime feeds the QAG engine with five first-class signals:
- Relevance — which rules apply to a given context.
- Overlap — where rule conditions intersect.
- Conflict — where rules produce contradictory outcomes.
- Redundancy — duplicate or near-duplicate rules.
- Predicate extraction — the condition component of a rule.
Coverage, coherence, and topology signals exist internally and will ship in later releases.
Intended use
Q-Prime is intended for use by:
- Regulated-industry engineering teams embedding rules, policies, contracts, and case documents.
- Compliance and audit functions running continuous rule-to-rule conflict detection across versioned policy sets.
- Regulated-news and research desks synthesizing multiple sources where the sign of the claim matters.
- Risk and model-governance leaders deploying an embedding layer whose failure mode is explainable, not statistical.
- Agent builders implementing conflict-aware long-term memory and context engineering.
Q-Prime is not intended for:
- General-purpose open-web retrieval or low-stakes Q&A.
- Non-English corpora (current release is English-only).
- Autonomous decisions that materially affect an individual's legal rights, employment, housing, credit, healthcare access, education, or liberty — except under a certified pipeline that includes qualified human review.
See License §5 (Responsible Use) for the binding language.
How to access
Q-Prime is distributed exclusively as a managed API. There are three access paths.
1. Evaluation access (researchers, engineers, non-production)
Click Request access above, or email contact@qgi.dev. Evaluation is governed by LICENSE.md §3 (90-day grant, non-production only, academic use permitted). Target SLA: 3 business days.
2. OpenRouter (pay-per-call)
Q-Prime will be listed on OpenRouter as part of the QAG engine progressive beta. Listing status and pricing are announced at qgi.dev.
3. Enterprise (production, SLA, audit, dedicated endpoints)
contact@qgi.dev. Tiers: Startup, Growth, Enterprise, OEM / Channel.
General availability of the full QAG engine is targeted for 21 June 2026. Q-Prime is available to selected customers in progressive beta before that date.
Quickstart (after you have an API key)
# pip install requests
import os, requests
API_URL = "https://api.qgi.dev/v1/qprime/embed"
API_KEY = os.environ["QGI_API_KEY"]
resp = requests.post(
API_URL,
headers={"Authorization": f"Bearer {API_KEY}"},
json={
"inputs": [
"A regulated entity must report any incident within 72 hours.",
"A regulated entity must not report an incident if law-enforcement investigation is active.",
],
"tasks": ["relevance", "conflict", "predicate"],
},
timeout=30,
)
resp.raise_for_status()
out = resp.json()
# out["conflict"][0][1] > 0 → the two clauses conflict on polarity + scope
A live demo is at QGI-dev/q-prime-demo.
What you do not get
To set expectations before first contact:
- No weights download. Q-Prime weights, adapters, and supporting parameters are not distributed. Redistribution requires a separately negotiated license.
- No training recipe. Data curation, training procedure, and internal evaluation methodology are confidential trade secrets (LICENSE.md §2.2).
- No architecture disclosure beyond the papers. Architectural details ship per the release cadence in the QAG paper series.
We are aware this is unusual for a model card on Hugging Face. Q-Prime is a commercial product, not an open research artifact. This card exists so developers and procurement teams can evaluate fit, not so the model can be cloned.
Evaluation
A headline result accompanies this card: on QGI's regulatory-conflict benchmark, Q-Prime lifts rule-conflict F1 from 0.000 (with a leading general-purpose embedding) to near-perfect on in-distribution data. Full evaluation methodology, out-of-domain results, multiple-backbone comparisons, throughput/latency numbers, and the full benchmark suite are released under evaluation agreement — contact@qgi.dev — and published in the forthcoming evaluation paper (Paper G in the QAG series).
Numbers on this card will be updated once Paper G is public.
Accompanying papers (QAG series)
| Paper | Focus |
|---|---|
| Paper A — Quantum-Augmented Generation (QAG) | Canonical engine paper. Core formalism, signals, architecture. |
| Paper B — Purpose-Built Embedding Models for Rule-Bearing Text | Position paper motivating Q-Prime. |
| Paper C — Conflict-Aware Memory for AI Agents | Long-term memory using QAG primitives. |
| Paper D — A Born-Rule Classifier | Zero-shot, calibrated categorization method note. |
| Paper E — Quantum HyperGraph (QHG) | First-class data model for rule-bearing knowledge. |
| Paper F — Beyond Retrieval-Augmented Generation | 2026 landscape review. |
| Paper G — Empirical Evaluation | Held under evaluation agreement; release with this card's numbers update. |
Each arXiv ID is added to this card's arxiv: YAML as it publishes.
Citation
@misc{sammane2026qprime,
title = {Q-Prime: A Quantum-Structured Embedding Model for Regulated AI},
author = {Sammane, Sam and {Quantum General Intelligence, Inc.}},
year = {2026},
howpublished = {Model card, \url{https://huggingface.co/QGI-dev/q-prime}},
note = {QAG engine documentation; accompanying paper series on arXiv}
}
Academic work using Q-Prime must cite per LICENSE.md §8.
Responsible use
Q-Prime may not be used to:
- Automate safety-critical decisions (as enumerated above) without qualified human review.
- Circumvent legal or regulatory obligations in the user's jurisdiction.
- Misrepresent the user's compliance posture to a regulator, counter-party, or auditor.
- Train a model intended to compete with Q-Prime, QAG, Neural Symbolic Agents, or any Qualtron model.
See LICENSE.md §5 for the binding language.
Contact
| Need | Where |
|---|---|
| Evaluation access, API keys, documentation | sam@qgi.dev |
| Commercial license, enterprise pilots, SLA, support | sam@qgi.dev |
| QAG engine waitlist (GA 21 June 2026) | qgi.dev |
| Partnership (cloud providers, hyperscalers, channel) | partner@qgi.dev |
| Press and analyst relations | press@qgi.dev |
| Security disclosure | security@qgi.dev |
Company
Quantum General Intelligence, Inc. — Delaware corporation, founded 2025.
Website: qgi.dev · Hugging Face: QGI-dev · GitHub: Quantum-General-Intelligence
© 2025–2026 Quantum General Intelligence, Inc. All rights reserved. "Q-Prime", "QAG", "Quantum-Augmented Generation", "QGI", "Neural Symbolic Agents", and "Qualtron" are trademarks of Quantum General Intelligence, Inc.