Mistral · released 2024-05 · 7.25B parameters
No sliding window in v0.3 (sliding_window null).
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.50 GB | 0.25 GB |
| 8,192 tokens | 1.00 GB | 0.50 GB |
| 32,768 tokens | 4.00 GB | 2.00 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 13.5 GB |
| FP8 (W8A8) | 8 | 7.25 GB |
| INT8 (W8A8 / W8A16) | 8 | 7.25 GB |
| NVFP4 | 4.5 | 4.51 GB |
| INT4 AWQ / GPTQ | 4.25 | 4.32 GB |
| GGUF Q8_0 | 8.5 | 7.17 GB |
| GGUF Q6_K | 6.56 | 5.54 GB |
| GGUF Q5_K_M | 5.69 | 4.80 GB |
| GGUF Q4_K_M | 4.89 | 4.13 GB |
| GGUF Q3_K_M | 3.91 | 3.30 GB |
| GGUF Q2_K | 3.35 | 2.83 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× · 325 tok/s | 1× · 146 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 231 tok/s | 1× · 94.0 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 175 tok/s | 1× · 67.9 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 248 tok/s | 1× · 103 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 172 tok/s | 1× · 48.5 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 53.5 tok/s | 1× · 18.6 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 93.8 tok/s | 1× · 33.8 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 124 tok/s | 1× · 36.0 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 80.8 tok/s | 2× · 36.6 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 56.4 tok/s | 1× · 11.9 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 17.7 tok/s | 1× · 6.0 tok/s |