Access CMMC Expert 12B (Research Model)
Memoriant, Inc. gates this research model with auto-approval. Login and contact sharing are required, but access is granted automatically upon request.
By requesting access you acknowledge:
- This is a research-tier model, NOT the Memoriant flagship.
- You will always review AI output before using it for compliance work.
- You will not submit AI-generated compliance documentation without qualified human review.
- Defense contractor compliance is your responsibility, not your AI tool's.
Acknowledge the responsible use terms below to access this research model.
Log in or Sign Up to review the conditions and access this model content.
CMMC Expert 12B (Research Model)
Version: 2026-q2 Tier: Research model (public preview of the Memoriant methodology) Valid through: June 30, 2026 Next release: July 1, 2026 (Q3 2026) License: Apache 2.0 (adapter) / inherits base-model license Publisher: Memoriant, Inc.
A 12B-parameter research model fine-tuned for cybersecurity compliance, specializing in CMMC 2.0, NIST SP 800-171, NIST SP 800-172, DFARS, HIPAA, and related regulatory frameworks. Built by Memoriant, Inc. — we build AI systems for regulated industries.
⚠️ This Is a Research Model — Not the Flagship
The Memoriant flagship compliance model is private and not distributed publicly.
This 12B research model is a public preview of the Memoriant methodology — it is published so that researchers, compliance teams, and AI builders can evaluate the approach, understand what a fine-tuned compliance model looks like, and benchmark their own training pipelines. It is not the model Memoriant deploys to enterprise customers.
| Tier | Public? | Where |
|---|---|---|
| Research Model (this model) | ✅ Yes — auto-gated | HuggingFace memoriant/cmmc-expert-12b |
| Specialized Models (Coming Q2 2026) | ✅ Public at launch | memoriant/cmmc-expert-nemotron-3-nano-4b (Lookup Specialist) and memoriant/cmmc-expert-nemotron-3-nano-30b (POA&M Generator) |
| Memoriant Flagship | ❌ Private | Enterprise deployment only — contact memoriant.ai |
If you are deploying compliance AI in production, this research model is not sufficient on its own. Contact Memoriant for commercial licensing, the flagship model, the enterprise platform, or custom deployment.
⚠️ AI Safety Disclaimer — Always Review Output
AI systems make mistakes. Always review AI-generated output before using it for any purpose.
Any AI system — including this one — can produce:
- Factually incorrect information — even with high benchmark scores
- Hallucinated citations — references to regulations, controls, or documents that do not exist
- Outdated guidance — AI knowledge reflects training cutoff, not current regulations
- Confident errors — AI often states wrong information with the same confidence as correct information
- Plausible-sounding fabrications — responses that read like expert advice but are invented
Before using any AI output for:
- Compliance documentation (SSPs, POA&Ms, audit responses)
- Regulatory submissions to DoD, NIST, or other agencies
- Assessment preparation or C3PAO engagements
- Internal policy or procedure creation
- Legal or contractual decisions
You must:
- Have a qualified human review every output
- Verify citations independently against authoritative sources
- Cross-check against NIST publications, DoD guidance, Federal Register, CMMC Assessment Guides
- Document the review process for audit purposes
- Never submit AI output directly — AI drafts are starting points, not finished products
This is especially critical for CMMC and defense compliance. Wrong answers can cause failed assessments. Failed assessments can cost DoD contracts. C3PAO assessors verify human understanding, not AI output. The DoD holds contractors accountable for their submissions, not the tools they used.
Memoriant's position: AI is a force multiplier for compliance professionals, not a replacement. The human stays accountable. The AI accelerates the work. Never let AI make final compliance decisions.
Model Details
| Detail | Value |
|---|---|
| Parameters | 12B |
| Quantization | Q5_K_M (GGUF) |
| Context Window | 128K tokens |
| Training Method | LoRA-based supervised fine-tuning |
| Eval Loss | 0.583 |
| Format | GGUF (compatible with Ollama, llama.cpp, LM Studio) |
| License | Apache 2.0 (adapter) |
Capability Categories
This research model is trained for the following general compliance task categories:
- Compliance Q&A — answering questions about CMMC 2.0, NIST SP 800-171 Rev 2/3, NIST SP 800-172, HIPAA, FedRAMP, and DFARS clauses
- Regulatory lookup — control-family and control-ID reference, framework definitions, terminology
- Document drafting assist — drafting compliance narrative starting points for human review
- Cross-framework reference — basic mapping between related federal and industry frameworks
- Hallucination resistance — trained to refuse fabricated constructs (e.g., "CMMC Level 4")
This list describes general categories. The specific training examples, task distribution, curation rules, adversarial scenarios, and reasoning patterns used to train Memoriant models are part of the proprietary training formula and are not disclosed.
Proprietary Training Formula
The specific data mixture, curation decisions, fine-tuning hyperparameters, adversarial example construction, deduplication strategy, and reasoning-style examples used to train Memoriant compliance models are proprietary trade secrets of Memoriant, Inc. and are not distributed with this release.
This research model was trained on a curated mix of NIST, DoD, and industry compliance publications. The exact mixture, ratios, and training recipe are proprietary.
This is deliberate. Regulatory fine-tuning is a valuable, differentiating asset — the specific formulation is what Memoriant sells. Publishing the full recipe would destroy both its trade secret status and its economic value. Publishing nothing would make the research non-reproducible. This model is the balance: a working research artifact that demonstrates the methodology, with the proprietary formulation held back for commercial use.
If you need the full Memoriant corpus, the flagship model, or custom training, contact Memoriant directly.
Intended Use
Good uses
- Research and evaluation of domain-specific compliance LLMs
- Benchmarking fine-tuning pipelines against a known-good reference model
- Educational use in AI safety and compliance contexts
- Academic research on narrow-task domain expert models
- Early-stage internal prototyping of compliance AI workflows
Not intended for
- Production compliance documentation without qualified human review
- Direct submission of AI output to DoD, NIST, or regulatory agencies
- Replacing qualified C3PAO assessments
- Classified or Sensitive Compartmented Information handling
- Use as the sole source of compliance guidance
- Any use case requiring guaranteed accuracy — this is a research model, not a validated production system
⚠️ Version Expiration
Valid through: June 30, 2026 Next release: July 1, 2026 (Q3 2026)
This model is dated. CMMC regulations, DFARS clauses, and NIST publications update continuously. Using a frozen model after its refresh date produces stale outputs that may have been accurate at training time but no longer reflect current regulations.
If you are using this model after the expiration date, your results are incomplete. The Memoriant model family is refreshed quarterly to incorporate new DFARS clauses, NIST SP 800-171/172 revisions, CMMC Program Office guidance, and assessment methodology updates.
A model trained on Q2 2026 data and deployed in Q4 2026 is already 6 months behind the regulatory landscape. For defense contractors, that 6-month gap can mean failed assessments, lost contracts, or documentation based on outdated requirements.
To get the latest version: Follow memoriant on HuggingFace, or subscribe to release notifications at memoriant.ai.
Quick Start
# Download and run with Ollama (requires HuggingFace login and auto-approved access)
ollama run memoriant/cmmc-expert-12b
# Or with llama.cpp
llama-server -m cmmc-expert-12b-q5_k_m.gguf --host 0.0.0.0 --port 8080
Recommended deployment: Air-gapped, local inference via Ollama or llama.cpp. No cloud dependency required.
Remember: Always review AI output before using it for compliance work. See the AI Safety Disclaimer above.
Limitations
- Research-tier only — this is not the Memoriant flagship. Quality ceiling is lower than the private flagship model.
- Knowledge cutoff — model knowledge reflects training cutoff, not real-time regulations
- Not a legal or assessment authority — responses must be verified against authoritative sources
- Not suitable for unreviewed compliance documentation — every output requires qualified human review
- Performance varies on organization-specific scenarios without retrieval-augmented generation (RAG) integration
- 12B parameter ceiling — smaller than the flagship; less detailed multi-step reasoning than 31B+ models
Related Memoriant Assets
Training data
- memoriant/cmmc-training-data-2026-q2 — Curated public subset of the Memoriant training corpus (methodology demonstration, not the full proprietary corpus)
Benchmarks
- memoriant/cmmc-benchmark-v1-preview-2026-q2 — 46-question preview (methodology sampler, not for validation)
- memoriant/cmmc-benchmark-v2-spotcheck-2026-q2 — 454-question spot check (triage only, not comprehensive)
- memoriant/cmmc-benchmark-v3-comprehensive-2026-q2 — 1,273-question comprehensive evaluation (the authoritative standard)
Coming Q2 2026 — Specialized models (public)
memoriant/cmmc-expert-nemotron-3-nano-4b— Edge tier (Lookup Specialist)memoriant/cmmc-expert-nemotron-3-nano-30b— Professional tier (POA&M Generator)
The Memoriant flagship model is not publicly available.
Changelog
2026-q2 (Current)
- Repositioned as research model — flagship is private
- Added auto-gated access with responsible use acknowledgement
- Added AI safety disclaimer
- Added expiration and refresh messaging
- Removed specific training-data mixture ratios (now proprietary)
- Removed specific fine-tuning hyperparameters (now proprietary)
- Added proprietary training formula language
- Updated related-asset links to quarterly-versioned names
Future Releases
- 2026-q3 (July 1, 2026)
- 2026-q4 (October 1, 2026)
About Memoriant
Memoriant, Inc. builds purpose-built AI systems for regulated industries. Our platform transforms domain expertise into deployable AI solutions that run on customer hardware or in authorized government-cloud enclaves — with the human always in the review loop.
Published by Memoriant, Inc. — A research-tier preview of our compliance AI methodology. The full Memoriant training formula and flagship model are proprietary.
- Downloads last month
- -
5-bit