Llama.
Open-weights model family — the default base for enterprise fine-tuning.
What it is.
Meta's open-weights model family. Llama 4 introduced Mixture-of-Experts and 10M-token context on Scout. Distributed under a near-permissive licence with a usage threshold above 700M monthly active users. The de-facto base model for enterprise fine-tuning.
Where it fits.
Buyers who need on-prem, air-gapped, or sovereign deployment. Also the procurement tool used to negotiate frontier-closed pricing — Llama on AWS Bedrock as the BATNA in every Anthropic and OpenAI Enterprise contract since 2025.
- Open weights enable on-prem and sovereign deployment
- Strong fine-tuning ecosystem and tooling
- Largest serving footprint across Bedrock, Azure, and Vertex
- Behind frontier-closed on reasoning benchmarks
- Self-serving needs GPU capacity most buyers do not own
Frequently asked.
What is the Llama licence?
Custom permissive licence. Free commercial use below 700M MAU. Above that, contact Meta. Most enterprises are clear.
Can I fine-tune Llama on customer data?
Yes. The standard pattern is LoRA or QLoRA fine-tunes served on Bedrock, Azure, or self-hosted GPU clusters.
Why fine-tune Llama instead of using a frontier-closed model?
Three reasons: data residency, cost per token at scale, and IP ownership of the resulting weights.
