Our catalog tracks 6 Llama models (1.2B–405.9 parameters). For each, the table shows the cheapest retail GPU that fully fits it at Q4 / 8K context and the fastest retail GPU, with decode tok/s. Fit means weights + KV cache + 1 GB runtime overhead within 95% of VRAM — not weights alone. 1 of them exceed every retail GPU we track and are marked accordingly. Methodology.
| # | Model | Params | Q4/8K needs (GB) | Best value GPU (fits) | Fastest retail GPU | Buy |
|---|---|---|---|---|---|---|
| 1 | Llama-3.2-1B-Instruct | 1.2B | 1.9 | Arc B570 10GB ≈201.5 tok/s est. | RTX 5090 32GB ≈950.3 tok/s est. | Search Amazon (affiliate) |
| 2 | Llama-3.2-3B-Instruct | 3.2B | 3.4 | Arc B570 10GB ≈75.6 tok/s est. | RTX 5090 32GB ≈356.4 tok/s est. | Search Amazon (affiliate) |
| 3 | Llama-3.1-8B-Instruct | 8.0B | 7.0 | Arc B570 10GB ≈30.2 tok/s est. | RTX 5090 32GB ≈214.3 tok/s measured | Search Amazon (affiliate) |
| 4 | Llama-3.1-70B-Instruct | 70.6B | 53.83 | Mac mini M4 Pro 64GB ≈5.0 tok/s measured | Mac Studio M3 Ultra 96GB ≈14.1 tok/s measured | Search Amazon (affiliate) |
| 5 | Llama-3.3-70B-Instruct | 70.6B | 53.83 | Mac mini M4 Pro 64GB ≈2.5 tok/s est. | MacBook Pro M4 Max 128GB ≈11.0 tok/s measured | Search Amazon (affiliate) |
| 6 | Llama-3.1-405B-Instruct | 405.9B | 305.25 | No retail GPU fits | — | — |