Llama 3.2 · released 2024-09 · 3.21B parameters
Memory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.44 GB | 0.22 GB |
| 8,192 tokens | 0.88 GB | 0.44 GB |
| 32,768 tokens | 3.50 GB | 1.75 GB |
| 131,072 tokens | 14.0 GB | 7.00 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 5.98 GB |
| FP8 (W8A8) | 8 | 3.36 GB |
| INT8 (W8A8 / W8A16) | 8 | 3.36 GB |
| NVFP4 | 4.5 | 2.21 GB |
| INT4 AWQ / GPTQ | 4.25 | 2.13 GB |
| GGUF Q8_0 | 8.5 | 3.18 GB |
| GGUF Q6_K | 6.56 | 2.45 GB |
| GGUF Q5_K_M | 5.69 | 2.13 GB |
| GGUF Q4_K_M | 4.89 | 1.83 GB |
| GGUF Q3_K_M | 3.91 | 1.46 GB |
| GGUF Q2_K | 3.35 | 1.25 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× · 485 tok/s | 1× · 186 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 373 tok/s | 1× · 122 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 299 tok/s | 1× · 88.9 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 395 tok/s | 1× · 133 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 278 tok/s | 1× · 60.3 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 103 tok/s | 1× · 24.8 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 173 tok/s | 1× · 44.7 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 197 tok/s | 1× · 44.7 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 138 tok/s | 1× · 26.4 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 101 tok/s | 1× · 17.5 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 35.4 tok/s | 1× · 8.0 tok/s |