Llama

Meta's Llama family is the reference point for local LLMs. Almost every GGUF build, inference engine, and fine-tune in the community supports Llama first, so "what GPU do I need for Llama?" is usually the first question a new buyer asks. The line covered here spans five sizes: Llama 3.2 3B, Llama 3.1 8B, Llama 3.1 70B, Llama 3.3 70B, and the Llama 4 Scout 109B MoE with 17B active parameters. All of them use Meta's community licenses (llama3.1 / llama3.2 / llama3.3 / llama4-community) rather than OSI licenses, and all ship with a 128K context window — except Scout, which Meta claims at 10M tokens.

The VRAM figures below use our standard estimate: Q4 weights + KV cache at 8K context + 1GB overhead. Longer contexts add roughly 0.025 GB per active billion parameters per 1K tokens, so 32K context costs noticeably more than 8K on the dense 70B models.

Size ladder

• Llama-3.2-3B-Instruct (3.2B dense) — Q4 needs only about 3.4 GB at 8K context. This runs on virtually any modern GPU, including older 8GB cards, with huge headroom for long context. A 12GB card is overkill — in a good way.
• Llama-3.1-8B-Instruct (8B dense) — Q4 lands around 7.0 GB at 8K. The sweet spot for 12GB cards like the RTX 3060 12GB or RTX 4070 class: fast, comfortable, and still leaves room for a 32K+ context. FP16 weights alone are about 16 GB, so full precision wants a 24GB card.
• Llama-3.1-70B-Instruct (70.6B dense) — Q4 needs about 53.8 GB at 8K, which means a single 48GB card is just short. The realistic Q4 target is a 96GB setup (RTX 6000 Pro 96GB, or 2×48GB). On 48GB you can drop to Q2 (~34.8 GB total) at a real quality cost. FP16 is a datacenter-only ~141 GB.
• Llama-3.3-70B-Instruct (70.6B dense) — identical size to 3.1 70B, so the same ~53.8 GB Q4 requirement and the same 96GB recommendation. Meta positioned 3.3 as delivering near-405B-class quality at 70B cost, which is why it has largely replaced 3.1 70B as the default big-Llama pick.
• Llama-4-Scout-17B-16E-Instruct (109B total, 17B active, 16 experts) — the family's MoE entry. Q4 weights are ~60 GB, but because only 17B parameters are active per token, KV cache at 8K adds only ~3.4 GB, for a total around 64.4 GB. That is a firm 96GB-tier model — but you get far better tok/s than a dense 109B would, closer to a 17B dense model's speed once loaded.

What makes the family special

The Llama line's real advantage is ecosystem gravity: every quant format, every runner (llama.cpp, Ollama, vLLM, LM Studio), and most fine-tunes treat Llama as the baseline. Context is uniformly 128K across the 3.x models, which is generous for local RAG — just remember KV cache scales with context, so a 70B at 128K needs many more GB than the 8K figures above. Scout is the outlier: a 10M-token claimed context (vendor number — treat long-context claims cautiously) and MoE efficiency that makes a 100B-class model plausible on a single high-end workstation card. Licensing is permissive for most commercial use under the community licenses, but read the terms if you are at very large scale.

Which one should you pick

• 12GB VRAM — Llama 3.2 3B or Llama 3.1 8B at Q4. The 8B is the default choice for quality; the 3B is for maximum speed or tiny form factors.
• 16GB VRAM — 8B at Q4 with lots of context headroom, or 8B FP16 if you insist on full precision (tight).
• 24GB VRAM — the 8B class is trivial here; you are really shopping for headroom, long context, or other 13B–30B models rather than Llama specifically.
• 48GB VRAM — 70B only at aggressive Q2 quants; for comfortable Q4 Llama 70B, go up a tier.
• 96GB VRAM — the Llama tier: 3.3 70B at Q4, or Llama 4 Scout with its MoE speed. This is where Llama stops being a compromise.
• 192GB+ — run Scout or 70B at FP16-class precision, or 70B with a genuinely long context window.

If you are unsure, start with the 8B on a 12–16GB card: it is the cheapest way to learn what your workload actually needs before spending on 96GB.

Links

• Best models for 12GB GPUs
• Best models for 16GB GPUs
• Best GPU for 70B models
• How much VRAM do local LLMs need?
• LLM quantization explained