Llama 3.1 8B Instruct in Meta’s official BF16 format needs 16.1 to 16.8 GiB at 8K context: about 15 GiB of weights plus 1 GiB of KV cache. That fits a 24 GB card like the RTX 4090, but not a 16 GB one. Llama 3.2 1B needs about 3 GiB and Llama 3.2 3B about 7.5 GiB. Llama 3.3 70B in BF16 needs 134 to 137 GiB, more than any single card we list.
All of these are the full-precision files Meta publishes. A 4-bit community GGUF would be much smaller, but we don’t list one, so there’s no Q4 row here.
Llama 3 in BF16 at 8K and 32K context
| Model | File size (bytes) | Weights (GiB) | KV at 8K (GiB) | Total at 8K (GiB) | KV at 32K (GiB) | Total at 32K (GiB) |
|---|---|---|---|---|---|---|
| Llama 3.2 1B Instruct | 2,471,645,608 | 2.30 | 0.25 | 2.68 to 3.12 | 1.00 | 3.43 to 3.92 |
| Llama 3.2 3B Instruct | 6,425,529,048 | 5.98 | 0.88 | 6.98 to 7.54 | 3.50 | 9.61 to 10.35 |
| Llama 3.1 8B Instruct | 16,060,556,376 | 14.96 | 1.00 | 16.08 to 16.83 | 4.00 | 19.08 to 20.04 |
| Llama 3.3 70B Instruct | 141,107,497,872 | 131.42 | 2.50 | 134.04 to 137.22 | 10.00 | 141.54 to 145.24 |
Official BF16 safetensors, one sequence, FP16 KV cache, llama.cpp. File sizes are the safetensors sizes in Meta’s Hugging Face listings (the repos are gated). Layer and head counts come from Meta’s llama-models repository (sku_list.py, commit 0e0b8c5).
KV cost per token: 32 KiB for the 1B, 112 KiB for the 3B, 128 KiB for the 8B and 320 KiB for the 70B. Going from 8K to 32K quadruples the cache; the weights don’t move.
Which Llama 3 for your card
- 8 GB: Llama 3.2 1B fits easily. The 3B in BF16 (6.98 to 7.54 GiB) is right at the 7 GiB an 8 GB card has free, so expect it to be tight or not fit.
- 12 GB: the 3B in BF16 runs comfortably at 8K and gets tight at 32K. The 8B in BF16 doesn’t fit.
- 16 GB: the 8B in BF16 still doesn’t fit; its weights alone are 14.96 GiB. You’d need a 4-bit conversion.
- 24 GB: the 8B in BF16 fits at 8K and is tight at 32K. At 128K it needs 31.08 to 32.88 GiB and doesn’t fit.
- 70B: the BF16 weights (131.42 GiB) are over an H100 80GB (about 78 GiB usable) and over a DGX Spark (112 GiB usable). Two 24 GB cards are not a 48 GB pool; see two GPUs and unified memory.
The 8B’s context curve is in context length VRAM tradeoff, and what fits on 12, 16 and 24 GB compares Llama with Qwen3, Gemma 3 and others at each size.
Model pages: Llama 3.2 1B, Llama 3.2 3B, Llama 3.1 8B, Llama 3.3 70B. For the general method, see how much VRAM a local LLM needs.
Numbers are estimates from model configs and published file sizes, not measurements. How we calculate.
FAQ
How much VRAM does Llama 3.1 8B need?
In Meta's official BF16 format, 16.08 to 16.83 GiB at 8K context. It fits a 24 GB card but not a 16 GB one. At 32K it rises to 19.08 to 20.04 GiB, which is tight on 24 GB.
Why is there no Q4 row for Llama?
Meta publishes Llama 3.x as BF16 safetensors, and we don't list community GGUF conversions for it. Rather than guess a file size, we only show the official files.
Does Llama 3.3 70B fit on one 80 GB GPU?
Not in BF16. The weights alone are 131.42 GiB, and the 8K total is 134.04 to 137.22 GiB, against about 78 GiB usable on an H100 80GB. Two separate GPUs don't add up into one memory pool either.
Which Llama 3 runs on an 8 GB GPU?
Llama 3.2 1B (2.68 to 3.12 GiB at 8K) fits easily. Llama 3.2 3B in BF16 needs 6.98 to 7.54 GiB, which is right at the edge of the 7 GiB an 8 GB card has free.
Numbers from calculator 0.3.0, model data 2026-10-02.1. Sources last checked 2026-10-29.