Llama: name the runtime, name the window.

Llama is the open-weight house. Llama Serve tells you how the model is hosted. Fine-tunes stay with Weights. Phone builds stay with Edge. An NVIDIA container stays with NIM.

Llama
Llama 3.3 · Llama 3.1
Open weights, community runtimes, 70B-class chat
Llama Serve

The weights that left the lab

Meta released Llama in February 2023 as a research download. Llama 2 made commercial use legitimate. Llama 3, then 3.1 and 3.3, turned a 70B-class open-weight model into the default starting point for a thousand companies. Groq, Together, and a long list of community runtimes taught people to serve those weights at conversational speed. Wrooms seats that lineage as Llama Serve: the question is not “what is Llama?” The question is how you host it.

Llama 3.3 and Llama 3.1

Llama 3.3 70B is the current deep default. Llama 3.1 70B remains on the card. Both are versatile chat models. Hermes on Nous is a different job — a tool-call trace, not a serve plan. Phi is a small Microsoft model when Windows is the point. NIM is an NVIDIA container. Three pages, three artifacts.

What Llama Serve will name

The runtime and the context window. It will not write a fine-tune recipe. It will not pick a phone quant. It will not pretend to be Meta’s download page. Lineage is flavor. You are still in a Wrooms room.

How to ask for a serve plan

Say the model and the context you need. @llamacpp or pin Llama Serve. Keep the room on host, not on a training run. Tinker is the training-run note. Weights is the adapter cookbook.