Llama 3.3 · released 2024-12 · 70.6B parameters
Memory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 1.25 GB | 0.63 GB |
| 8,192 tokens | 2.50 GB | 1.25 GB |
| 32,768 tokens | 10.0 GB | 5.00 GB |
| 131,072 tokens | 40.0 GB | 20.0 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 131.4 GB |
| FP8 (W8A8) | 8 | 69.6 GB |
| INT8 (W8A8 / W8A16) | 8 | 69.6 GB |
| NVFP4 | 4.5 | 42.6 GB |
| INT4 AWQ / GPTQ | 4.25 | 40.7 GB |
| GGUF Q8_0 | 8.5 | 69.8 GB |
| GGUF Q6_K | 6.56 | 53.9 GB |
| GGUF Q5_K_M | 5.69 | 46.7 GB |
| GGUF Q4_K_M | 4.89 | 40.2 GB |
| GGUF Q3_K_M | 3.91 | 32.1 GB |
| GGUF Q2_K | 3.35 | 27.5 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× · 52.5 tok/s | 1× · 35.1 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 32.7 tok/s | 1× · 21.6 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 2× · 41.6 tok/s | 2× · 28.2 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 35.9 tok/s | 1× · 23.7 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 23.9 tok/s | 2× · 23.7 tok/s |
| L40S 48GBFP8 (W8A8) | 2× · 11.7 tok/s | 4× · 12.3 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 11.3 tok/s | 2× · 13.4 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 2× · 30.6 tok/s | 4× · 23.1 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 4× · 26.5 tok/s | 8× · 16.6 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 7.1 tok/s | 1× · 1.4 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 2.0 tok/s | > 16 |