Concept card

Meta Llama — Open-Weight Frontier Models for Production Adoption

As of 2026-10-10

Edited by Cédric Mary, sole editor and publication director — 27 years in the SAP ecosystem. How these pages are produced, and where automation is used.

What is Meta Llama?

Llama's strategic point isn't capability parity with closed models — it's that the weights download and run on the customer's own GPUs, which is what matters for EU residency and fine-tuning cost.

What it is

Meta Llama is the open-weight frontier model family from Meta AI, distinguishing itself from Claude / GPT / Gemini by publishing the model weights under a community licence that permits production use (subject to size-of-business and acceptable-use clauses). As of May 2026 the Llama line is the dominant open-weight option for enterprises that need self-hosted or on-premise inference, with Llama variants spanning multiple parameter sizes for different latency / cost / capability points. Specific current-generation version naming (e.g. Llama 4) evolves; ai.meta.com is the authoritative reference for the current release.

The open-weight posture is the strategic point. Where Claude / GPT-5 / Gemini are accessible only through vendor APIs (the weights stay behind the vendor's serving infrastructure), Llama weights download from Meta or Hugging Face and run on the customer's own GPUs — AWS p5 instances, Azure ND-H100, on-premise NVIDIA DGX clusters, or sovereign-cloud deployments in regulated jurisdictions. This architectural difference matters for three customer profiles: regulated EU customers needing strict data residency, customers fine-tuning on proprietary corpora at scale where API-based fine-tuning is cost-prohibitive, and customers requiring inference-cost predictability independent of vendor pricing decisions.

Why it matters

Three customer profiles specifically need this: regulated EU customers needing strict data residency, customers fine-tuning at scale where API-based fine-tuning is cost-prohibitive, and customers wanting cost predictability independent of vendor pricing.
Llama has historically tracked one generation behind closed frontier models on public benchmarks (3.x vs GPT-4o/Claude 3.5/Gemini 1.5); whether Llama 4 closes that gap is vendor-self-reported, not peer-reviewed.
Stanford HAI's AI Index 2026 documents Llama as the most-downloaded open-weight frontier family among enterprise users in 2025-2026, widely deployed via Databricks Mosaic AI, AWS Bedrock and Azure ML.

Key points

Strategic differentiator — Meta publishes Llama weights under community licence permitting production use (subject to size + acceptable-use clauses); Claude / GPT-5 / Gemini do NOT publish weights.
Deployment surfaces — Llama runs on customer GPUs (AWS p5, Azure ND-H100, on-premise NVIDIA DGX, sovereign-cloud) via Databricks Mosaic AI, AWS Bedrock, Azure ML, Hugging Face Inference Endpoints — Stanford HAI AI Index 2026 documents Llama as most-downloaded open-weight frontier family.
Three target customer profiles — regulated EU customers needing strict data residency, customers fine-tuning on proprietary corpora at scale, customers requiring inference-cost predictability independent of vendor pricing.
Capability tier (historical) — Llama has tracked one generation behind the frontier closed models (Llama 3.x vs GPT-4o / Claude 3.5 / Gemini 1.5); whether current Llama 4 generation closes the gap is vendor self-reported and peer-review pending.
Trade-offs — self-hosting requires GPU infrastructure expertise, model-serving operations, security hardening (no vendor SOC-2 / ISO-27001 inheriting); higher op-ex baseline but lower marginal cost per inference at scale.
SAP fit — Llama reaches Joule through SAP AI Agent Hub (C211); self-hosted Llama + Hub governance + observability is the emerging regulated-EU reference architecture.

Terms used on this page

Open-weight model: An AI model where the trained parameter weights are published and downloadable, allowing customers to run inference on their own infrastructure; contrast with closed-weight frontier models (Claude, GPT-5, Gemini) accessible only via vendor APIs.
Llama community licence: Meta's open-weight licence permitting commercial use of Llama models; the Llama 4 edition sets a concrete threshold — a deployer whose product exceeds 700 million monthly active users must obtain a separate licence directly from Meta (Hugging Face model card, checked 2026-09-27).
Self-hosted inference: Deployment pattern where the customer runs the AI model on its own GPU infrastructure (AWS p5, Azure ND-H100, on-premise NVIDIA DGX, sovereign cloud); strict data residency and cost predictability at the price of operational complexity.
Databricks Mosaic AI: Databricks' AI platform that hosts open-weight models (Llama prominent) with managed serving, fine-tuning and governance; common deployment surface for Llama in SAP customer environments via the OEM-inside-BDC Databricks connector.
Active vs total parameters (mixture-of-experts): In a mixture-of-experts model like Llama 4 Scout, the total parameter count (109B) is not what runs on a given token — only the active subset (17B, routed across 16 experts) is computed per forward pass; GPU sizing should follow the active figure, not the headline total.
SAP AI Agent Hub: SAP's cross-vendor system of record for agents, LLMs and MCP servers, discovering assets across Microsoft, Google, AWS, ServiceNow and SAP AI Core; the governance path for a Llama endpoint that generative AI hub does not natively host.

Sources

Full card available to members. What the full card adds: the full decision framework · the SAP vs Snowflake / Databricks / Fabric comparison · the common pitfalls and their fix · the cheat sheet · the architecture schemas · the code blocks · the facts worth quoting.

Open in the app →