Llama 3.3 70B Instruct
Meta · 70.6B parameters · 131,072 token max context · Model card
Memory needed
| Quant | Weights | Total @ 8K ctx |
|---|---|---|
| Q3_K_M | 34.4 GB | 38.9 GB |
| Q4_K_M | 42.8 GB | 47.6 GB |
| Q5_K_M | 50.3 GB | 55.3 GB |
| Q6_K | 58.2 GB | 63.5 GB |
| Q8_0 | 75.0 GB | 80.7 GB |
| F16 | 141 GB | 149 GB |
Approximate GGUF sizes from bits-per-weight; totals include fp16 KV cache and ~0.8 GB overhead.
Can my hardware run it?
| Device | Verdict | Best quant | Est. tok/s (Q4) |
|---|---|---|---|
| Apple M5 Max (40-core GPU, 128GB) | Yes, usable | Q8_0 | ~10.7 |
| Apple M2 Ultra (76-core GPU, 192GB) | Yes, usable | Q8_0 | ~8.4 |
| Apple M3 Ultra (80-core GPU, 512GB) | Yes, usable | F16 | ~8.2 |
| Apple M4 Max (40-core GPU, 128GB) | Yes, usable | Q8_0 | ~7.4 |
| Apple M5 Pro (20-core GPU, 64GB) | Barely | Q4_K_M | ~5.9 |
| Apple M3 Max (40-core GPU, 64GB) | Barely | Q4_K_M | ~5.9 |
| Apple M1 Max (32-core GPU, 64GB) | Barely | Q4_K_M | ~5.5 |
| NVIDIA DGX Spark | Barely | Q8_0 | ~5.1 |
| NVIDIA GeForce RTX 5090 | No | — | — |
| NVIDIA GeForce RTX 4090 | No | — | — |
| NVIDIA GeForce RTX 3090 | No | — | — |
| NVIDIA GeForce RTX 3090 Ti | No | — | — |
| NVIDIA GeForce RTX 5080 | No | — | — |
| NVIDIA GeForce RTX 5070 Ti | No | — | — |
| NVIDIA GeForce RTX 5070 | No | — | — |
| NVIDIA GeForce RTX 5060 Ti 16GB | No | — | — |
| NVIDIA GeForce RTX 4080 SUPER | No | — | — |
| NVIDIA GeForce RTX 4070 Ti SUPER | No | — | — |
| NVIDIA GeForce RTX 4060 Ti 8GB | No | — | — |
| NVIDIA GeForce RTX 3060 12GB | No | — | — |
| NVIDIA GeForce RTX 3080 10GB | No | — | — |
| AMD Radeon RX 7900 XTX | No | — | — |
| Apple M4 (10-core GPU, 24GB) | No | — | — |
| Apple M4 Pro (20-core GPU, 48GB) | No | — | — |
Speeds estimated from memory bandwidth (calibrated per device with a measured llama.cpp benchmark where available).