Llama 3.2 · released 2024-09 · 1.24B parameters
Memory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.13 GB | 0.06 GB |
| 8,192 tokens | 0.25 GB | 0.13 GB |
| 32,768 tokens | 1.00 GB | 0.50 GB |
| 131,072 tokens | 4.00 GB | 2.00 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 2.30 GB |
| FP8 (W8A8) | 8 | 1.40 GB |
| INT8 (W8A8 / W8A16) | 8 | 1.40 GB |
| NVFP4 | 4.5 | 1.00 GB |
| INT4 AWQ / GPTQ | 4.25 | 0.97 GB |
| GGUF Q8_0 | 8.5 | 1.22 GB |
| GGUF Q6_K | 6.56 | 0.94 GB |
| GGUF Q5_K_M | 5.69 | 0.82 GB |
| GGUF Q4_K_M | 4.89 | 0.70 GB |
| GGUF Q3_K_M | 3.91 | 0.56 GB |
| GGUF Q2_K | 3.35 | 0.48 GB |
One server, 8K context, FP8 on data center GPUs and GGUF Q4_K_M on consumer and Apple hardware.
| GPU | 1 user | 32 users |
|---|---|---|
| B200 SXM (HGX B200)FP8 (W8A8) | 1× · 734 tok/s | 1× · 433 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 621 tok/s | 1× · 314 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 532 tok/s | 1× · 242 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 644 tok/s | 1× · 336 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 490 tok/s | 1× · 174 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 226 tok/s | 1× · 75.8 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 351 tok/s | 1× · 132 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 335 tok/s | 1× · 126 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 266 tok/s | 1× · 79.8 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 212 tok/s | 1× · 54.9 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 84.3 tok/s | 1× · 25.3 tok/s |