Registry / Llama via Azure AI
Llama via Azure AI
Meta's Llama models served as third-party serverless (Models-as-a-Service / standard) deployments on Microsoft Azure AI Foundry. Microsoft hosts the model on Microsoft-managed infrastructure and acts as data processor; Meta remains the model developer and the GPAI provider for EU AI Act purposes.
Watch-outs 7
The cells where this offering is not a clean public yes. This is what to check before you sign.
- SOC 2 Type II: Partial SOC 2 covers Azure services but does not include the Llama model
- ISO 27001: Partial Azure's ISO/IEC 27001 certificate covers the Azure platform (and Azure AI services on it) but not Meta's Llama model itself, and Meta publishes no ISO 27001 certificate for Llama.
- ISO 42001: Partial Llama is not listed in Microsoft’s ISO 42001 scope; Meta has no public ISO 42001 certification.
- HIPAA BAA: Partial Microsoft’s BAA applies at the Azure service level, but does not explicitly cover Llama models
- Residency: Partial Standard (pay-per-token) deployments only support Global Standard — no data pinning. Only version 9 of Llama-3.3-70B offers data-zone-provisioned-managed (EU) via provisioned throughput (reserved capacity, not standard billing). Llama-4-Maverick has Global Standard only.
- GPAI Code: No, verified The official EU signatory list shows no entry for Meta, the developer of Llama, confirming its absence and justifying no_verified per the GPAI rule.
- Art. 53 summary: No public evidence no dedicated Article 53 public summary found on Meta’s site
Azure SOC 2 report lists only Azure, Dynamics 365, Microsoft 365, and Power Platform, so the Llama via Azure AI model is not in scope, yielding only partial coverage.
For a list of Microsoft cloud services in audit scope, see the Azure SOC 2 Type 2 attestation report... - Azure - Dynamics 365 - Microsoft 365 - Power Platform
The recorded source trust.anthropic.com is Anthropic's trust portal covering Claude, not Llama; the authoritative evidence is Microsoft's own page confirming Azure's ISO/IEC 27001, which certifies the serving platform but not the third-party Llama model (Meta has no public ISO 27001 covering Llama),
Microsoft Azure, Dynamics 365, and other Microsoft online services undergo regular independent third-party audits for ISO/IEC 27001 compliance. ... The Azure ISO/IEC 27001 certificate covers Azure, Dynamics 365, Power Platform, and select Microsoft 365 cloud services.
Microsoft’s ISO 42001 certification covers only the listed services and does not include Llama, and there is no public evidence that Meta (Llama’s developer) holds its own ISO 42001 certification.
## Microsoft AI services in scope for ISO 42001 certification - GitHub Copilot - Microsoft 365 Copilot - Microsoft Copilot Health - Microsoft Copilot Studio - Microsoft Dragon Copilot - Microsoft Dragon Copilot (Radiologist) - Microsoft Foundry - Microsoft Security Copilot
The Microsoft Service Trust Portal is publicly accessible and serves as the maintained compliance portal for Azure AI services, including Llama via Azure AI.
Learn how Microsoft cloud services protect your data, and how you can manage cloud data security and compliance for your organization.
The provider’s own Q&A explicitly states it cannot confirm model‑by‑model HIPAA coverage, so coverage for Llama is not publicly verified.
Based on the currently available Microsoft documentation, we **cannot confirm which specific AI models (for example, Anthropic Claude Sonnet, Claude Opus, or individual Azure OpenAI models) are covered for processing Protected Health Information (PHI) under the Microsoft General HIPAA Business Associate Agreement (BAA)** ... **does not identify specific Azure AI Foundry models or model providers as individually covered under the BAA**.
Microsoft's DPA is public on its own domain (aka.ms/DPA) and itself contains the EU/UK Standard Contractual Clauses, and the subprocessor list is also publicly published on Microsoft's trust center page (microsoft.com/en-us/trust-center/privacy/data-access), so both elements of the question ("DPA wi
Download the most recent version (English) (May 2026): Microsoft Product and Services Data Protection Addendum (WW); "The Microsoft Online Services Subprocessor List identifies subprocessors authorized to subprocess customer or personal data in Microsoft Online Services. This list is applicable for all Microsoft Online Services governed by the Microsoft Data Protection Addendum."
The provider's own public documentation explicitly states that prompts and outputs are not used to train or improve any models, confirming a public no‑training commitment for Llama on Azure AI.
Microsoft doesn't use these prompts and outputs to train or improve Microsoft models, the model provider's models, or any third party's models.
The cited Azure OpenAI "Models sold by Azure" page does not cover Llama (a third-party catalog model); the authoritative Microsoft page for Llama-on-Azure is the model-catalog data-privacy page (learn.microsoft.com/en-us/azure/foundry-classic/how-to/concept-data-privacy), which documents statelessne
Models are stateless, and they don't store any prompts or outputs. If content filtering is enabled, the Azure AI Content Safety service screens prompts and outputs for certain categories of harmful content in real time.
The recorded source describes general Azure AI Foundry platform capabilities, not Llama-specific support. Multiple sources confirm Llama models lack Data Zone Standard and Standard (single region) deployment types: Microsoft Q&A confirms users cannot select Data Zone Standard for Llama, only Global
Llama-3.3-70B-Instruct version 1-5: global-standard (only). Version 9: global-standard, global-provisioned-managed, data-zone-provisioned-managed (EU), data-zone-provisioned-managed (US)
The official EU signatory list shows no entry for Meta, the developer of Llama, confirming its absence and justifying no_verified per the GPAI rule.
AI Studio Delta - Aleph Alpha - Almawave - Amazon - Anthropic - Black Forest Labs - Bria AI - Cohere - Domyn - Dweve - Fastweb - IBM - LINAGORA - Microsoft - Mistral AI - Open Hippo - OpenAI - Pleias - ServiceNow - WRITER
Search of Meta’s official pages found only generic training‑data descriptions and no EU‑Article 53 style public summary for Llama.
"Meta ... includes a mix of publicly available data, licensed data and information from Meta’s products and services. Generally, models are trained on trillions of tokens, which are pieces of information of different types."
Spotted an error? Submit a correction with evidence, corrections with a primary source are folded in and credited in the changelog.