Mistral · released 2024-07 · 12.2B parameters
head_dim 128 (32 x 128 = 4096 < hidden 5120). Tekken tokenizer.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.63 GB | 0.31 GB |
| 8,192 tokens | 1.25 GB | 0.63 GB |
| 32,768 tokens | 5.00 GB | 2.50 GB |
| 131,072 tokens | 20.0 GB | 10.0 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 22.8 GB |
| FP8 (W8A8) | 8 | 13.9 GB |
| INT8 (W8A8 / W8A16) | 8 | 13.9 GB |
| NVFP4 | 4.5 | 10.0 GB |
| INT4 AWQ / GPTQ | 4.25 | 9.73 GB |
| GGUF Q8_0 | 8.5 | 12.1 GB |
| GGUF Q6_K | 6.56 | 9.35 GB |
| GGUF Q5_K_M | 5.69 | 8.11 GB |
| GGUF Q4_K_M | 4.89 | 6.97 GB |
| GGUF Q3_K_M | 3.91 | 5.58 GB |
| GGUF Q2_K | 3.35 | 4.78 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× · 207 tok/s | 1× · 105 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 139 tok/s | 1× · 66.4 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 103 tok/s | 1× · 47.5 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 151 tok/s | 1× · 72.6 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 89.3 tok/s | 1× · 33.6 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 29.3 tok/s | 1× · 12.8 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 52.5 tok/s | 1× · 23.4 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 84.4 tok/s | 2× · 45.1 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 52.7 tok/s | 2× · 27.8 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 36.1 tok/s | 1× · 7.5 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 9.5 tok/s | 1× · 4.1 tok/s |