Llama 3.1 · released 2024-07 · 406B parameters
8 KV heads (official layout; some early mirrors carry a 16-KV-head 'MP16' config). Official FP8 checkpoint also published.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 1.97 GB | 0.98 GB |
| 8,192 tokens | 3.94 GB | 1.97 GB |
| 32,768 tokens | 15.8 GB | 7.88 GB |
| 131,072 tokens | 63.0 GB | 31.5 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 756.0 GB |
| FP8 (W8A8) | 8 | 385.8 GB |
| INT8 (W8A8 / W8A16) | 8 | 385.8 GB |
| NVFP4 | 4.5 | 223.9 GB |
| INT4 AWQ / GPTQ | 4.25 | 212.3 GB |
| GGUF Q8_0 | 8.5 | 401.6 GB |
| GGUF Q6_K | 6.56 | 309.9 GB |
| GGUF Q5_K_M | 5.69 | 268.8 GB |
| GGUF Q4_K_M | 4.89 | 231.0 GB |
| GGUF Q3_K_M | 3.91 | 184.7 GB |
| GGUF Q2_K | 3.35 | 158.3 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) | 4× · 29.8 tok/s | 4× · 26.5 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 4× · 20.3 tok/s | 4× · 17.7 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 8× · 20.7 tok/s | 8× · 18.6 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 4× · 21.9 tok/s | 4× · 19.2 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 4× · 16.3 tok/s | 8× · 18.4 tok/s |
| L40S 48GBFP8 (W8A8) | 16× · 6.7 tok/s | 16× · 5.3 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 8× · 9.6 tok/s | 8× · 8.0 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 16× · 8.8 tok/s | 16× · 7.4 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 16× · 7.8 tok/s | > 16 |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | > 16 | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |