Meta Llama
Meta Platforms, Inc. · EFROS US AI Vendor Governance Index entry
Composite governance score
F = inadequate posture for any regulated workload. Re-evaluate before procurement.
About this vendor
Open-weight foundation model family (Llama 3.x, Llama 4) distributed under a community license. Used primarily as a self-hosted or partner-hosted alternative to API-only vendors.
- Enterprise tier
- Self-hosted (open weights) or cloud-hosted via Bedrock, Azure AI, Vertex AI, Together, Fireworks, Groq
- Consumer tier
- Meta AI consumer (meta.ai)
- Vendor homepage
- https://llama.com
Twelve-axis governance scoring
Each axis is scored Yes / Partial / No / N/A against public evidence — vendor trust portals, BAAs/DPAs, SOC 2 report cover pages, published methodology documents. N/A applies when the axis is structurally inapplicable (foundation models, for example, defer Section 1557 to the downstream healthcare deployer).
| Axis | Status | EFROS note | Source |
|---|---|---|---|
| BAA / DPA available | No | Meta does not offer a BAA directly. BAA must be obtained from the hosting partner (AWS Bedrock, Azure AI Studio, GCP Vertex) where Llama is deployed. Self-hosted deployments shift the entire BAA burden to the deploying organization. | Meta Llama Community License |
| Training-data opt-out | Yes | Open weights — no training feedback loop to Meta. Inputs to your hosted deployment never leave your tenant. | Meta Llama license terms |
| US data residency option | Yes | Self-hosted or partner-hosted on a US region — deploying organization controls residency entirely. | Deployment-controlled |
| SOC 2 Type II report | No | Meta does not provide SOC 2 for Llama directly. Hosting partner (AWS/Azure/GCP) provides cloud-side SOC 2. | Meta Trust Center |
| ISO/IEC 42001 attestation | No | No ISO/IEC 42001 attestation. | Public posture review |
| NIST AI RMF self-attestation | No | No NIST AI RMF self-attestation. Meta publishes Responsible Use Guide and Model Card; deploying organization performs RMF mapping. | Meta Responsible Use Guide |
| Colorado AI Act readiness | No | No Colorado AI Act compliance statement. Deployer responsibility entirely. | Public posture review |
| HHS-OCR Section 1557 readiness | N/A | Foundation model — Section 1557 is deployer responsibility. | HHS-OCR Section 1557 — deployer scope |
| FRB SR 11-7 readiness | N/A | Foundation model — SR 11-7 is deployer responsibility. | FRB SR 11-7 — deployer scope |
| ABA Formal Op 512 readiness | N/A | Foundation model — ABA Op 512 is deployer responsibility. | ABA Formal Op 512 — practitioner scope |
| Subprocessor list public | No | Self-hosted: no Meta subprocessor chain. Partner-hosted: hosting partner's subprocessor list applies. | Deployment-controlled |
Trust-center maturity
Meta publishes Responsible Use Guide, model cards, license terms. No trust portal in the OpenAI/Anthropic sense. Compliance posture lives at the hosting layer.
Source: llama.com
Deep dive
Overview
Llama scores poorly on a vendor-governance scorecard because Meta delegates governance to the deploying organization. This is by design — open weights mean the deployer owns the entire stack. The right way to evaluate Llama is to score the hosting partner (AWS Bedrock, Azure AI, Vertex AI) instead, because that's where the BAA, SOC 2, residency, and subprocessor controls actually live.
Strengths
- Open weights — full deployer control of data, residency, retention
- No training feedback loop to Meta
- Cost advantage at scale via self-hosting
Weaknesses
- No vendor-side BAA, SOC 2, residency, or subprocessor controls
- Deployer owns 100% of governance burden
- No NIST AI RMF self-attestation, no Colorado AI Act statement
Best-fit use case
Organizations with mature ML/AI platform teams that need full data control, are running on-prem or sovereign-cloud workloads, or have validated hosting on AWS Bedrock / Azure AI Studio / GCP Vertex with the hosting partner's BAA in place.
Avoid when
Smaller organizations without an internal AI platform team. The cost of building deployer-side governance on top of Llama exceeds the cost of paying for OpenAI Enterprise or Claude for Work in most mid-market scenarios.
Operator's take
Deploy Meta Llama when organizations with mature ML/AI platform teams that need full data control, are running on-prem or sovereign-cloud workloads, or have validated hosting on AWS Bedrock / Azure AI Studio / GCP Vertex with the hosting partner's BAA in place. The composite score of 25 (grade F) reflects a mixed posture for regulated US workloads. Skip the vendor when smaller organizations without an internal AI platform team. The cost of building deployer-side governance on top of Llama exceeds the cost of paying for OpenAI Enterprise or Claude for Work in most mid-market scenarios. In every deployment, treat the cells above as a snapshot — the acquisition that gets to production safely is the one that re-verifies the trust-center posture before contract signature and rebuilds the matrix at renewal.
How this scoring is computed
The composite score blends eleven scoreable axes (BAA, training opt-out, US data residency, SOC 2, ISO/IEC 42001, NIST AI RMF, Colorado AI Act, Section 1557, SR 11-7, ABA Op 512, subprocessor transparency) with the trust-center maturity score. Axes marked N/A are excluded from the denominator so vendors are not penalized for sector-inapplicable axes. The vendor's primary sector amplifies the most relevant axes — healthcare vendors weight Section 1557 ×2, legal vendors weight ABA Op 512 ×2, banking vendors weight SR 11-7 ×2 — so the composite reflects what matters in the actual buying context.
Read the full methodology →Disagree with this scoring?
EFROS publishes scoring rationale per cell with a public source. If you have evidence that a specific axis should score differently — a new BAA, a new certification, a documented policy change — submit a formal challenge below. We re-score and publish the result with the next quarterly edition (or as a mid-quarter changelog entry if the change is material).
Disagree with a score?
Every cell in the EFROS Index is source-cited. If you have a public source that contradicts a score for Meta Llama, submit a formal challenge — we re-verify against the source and respond within 14 days.
Other vendors in Foundation model
Same category, scored on the same twelve axes. Useful for head-to-head shortlisting.
Take the scoring into production
The Index tells you the posture. These engagements turn the posture into a deployable program — vendor selection, governance policy, sector overlay, audit-ready evidence.