Llama — what creators actually do with it

The Llama tag is thin and often incidental across its 7 posts. One really concerns llama.cpp, the runtime: @codrkai installs it with Homebrew, starts llama-server with a context size and port, and runs a 9B Qwen GGUF at 16-bit (about 17 GB) on a 32 GB MacBook Pro (llama.cpp setup).

The posts about Meta's model itself are older launch coverage. @only_agis_ recaps the Llama 3.1 release in 8B, 70B and 405B open weights, described as topping evals at the time, with a demo of iteratively editing a travel itinerary table (Llama 3.1). @mattfarmerai covers Nemotron 70B, NVIDIA's fine-tune of Llama 3.1 70B Instruct, which the creator says reached third on the LM Arena leaderboard (Nemotron 70B).

The more durable material uses Llama as a reference model in explainers. @adam_rosler notes that the KV cache at 32k tokens on Llama 70B takes 17 GB of GPU memory per request, then walks through a 3-bit polar quantization trick (KV cache), and a companion video explains rotary position embeddings (RoPE). A watermarking explainer also tags it, though its claims are unverified (watermarks).

DeepSeek co-occurs in 3 posts; Hugging Face, ChatGPT and Gemini in 2 each. One cost-cutting post lists Llama among router fallbacks after Anthropic's subscription change (router fallback). This page suits readers who want to understand what open-weight models cost to serve, more than a how-to.