Mixtral · released 2024-04 · 141B parameters (39.2B active per token)
Memory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.88 GB | 0.44 GB |
| 8,192 tokens | 1.75 GB | 0.88 GB |
| 32,768 tokens | 7.00 GB | 3.50 GB |
| 65,536 tokens | 14.0 GB | 7.00 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 261.9 GB |
| FP8 (W8A8) | 8 | 131.7 GB |
| INT8 (W8A8 / W8A16) | 8 | 131.7 GB |
| NVFP4 | 4.5 | 74.8 GB |
| INT4 AWQ / GPTQ | 4.25 | 70.7 GB |
| GGUF Q8_0 | 8.5 | 139.2 GB |
| GGUF Q6_K | 6.56 | 107.4 GB |
| GGUF Q5_K_M | 5.69 | 93.2 GB |
| GGUF Q4_K_M | 4.89 | 80.1 GB |
| GGUF Q3_K_M | 3.91 | 64.0 GB |
| GGUF Q2_K | 3.35 | 54.8 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× · 94.0 tok/s | 2× · 45.2 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 2× · 97.7 tok/s | 2× · 28.4 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 4× · 97.0 tok/s | 4× · 35.4 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 65.3 tok/s | 1× · 16.6 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 2× · 80.4 tok/s | 2× · 20.0 tok/s |
| L40S 48GBFP8 (W8A8) | 4× · 32.4 tok/s | 8× · 14.4 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 2× · 36.8 tok/s | 4× · 16.7 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 4× · 54.8 tok/s | 8× · 24.8 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 8× · 35.3 tok/s | 8× · 18.2 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 12.6 tok/s | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |