Meta’s Llama 3 70B has become the base model of choice for enterprise fine-tuning programs in a way Llama 2 never fully achieved. Three things changed: the base model quality is close enough to GPT-4 class on most enterprise tasks that the performance gap no longer justifies closed-model lock-in for fine-tuning candidates; the commercial license is genuinely permissive for most enterprise use cases; and the fine-tuning tooling ecosystem matured around it.
The downstream effects are worth tracking. First, managed fine-tuning services built on top of Llama 3 (Fireworks, Together AI, Modal) are seeing material volume shifts away from OpenAI’s fine-tuning API. The OpenAI fine-tuning product has better UX but worse economics at scale, and enterprises doing serious fine-tuning programs are increasingly sensitive to that gap.
Second, the open-source fine-tuning toolchain — LoRA via PEFT, quantization via bitsandbytes, serving via vLLM — has reached a point where a mid-size engineering org can build and deploy a fine-tuned Llama 3 variant without specialist ML infrastructure. Six months ago that required a small dedicated team. Now it’s a two-week project for engineers who can read documentation.
Third, the regulatory and compliance picture is forcing the issue. Regulated industries that cannot send data to OpenAI’s API but can run workloads on their own infrastructure are evaluating Llama 3 on its merits, and for many healthcare and financial services use cases, the fine-tuned Llama 3 variant is good enough to go to production.
The risk: Meta’s continued investment in the Llama line is strategic, not purely altruistic. If the competitive dynamics shift and Meta needs to pull back on commercial licensing, enterprises that built on Llama face a hard transition. That’s a tail risk worth pricing in.