Mistral Large · released 2024-11 · 123B parameters
Same architecture as Mistral-Large-Instruct-2407 (Jul 2024).
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 1.38 GB | 0.69 GB |
| 8,192 tokens | 2.75 GB | 1.38 GB |
| 32,768 tokens | 11.0 GB | 5.50 GB |
| 131,072 tokens | 44.0 GB | 22.0 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 228.4 GB |
| FP8 (W8A8) | 8 | 115.7 GB |
| INT8 (W8A8 / W8A16) | 8 | 115.7 GB |
| NVFP4 | 4.5 | 66.4 GB |
| INT4 AWQ / GPTQ | 4.25 | 62.9 GB |
| GGUF Q8_0 | 8.5 | 121.3 GB |
| GGUF Q6_K | 6.56 | 93.6 GB |
| GGUF Q5_K_M | 5.69 | 81.2 GB |
| GGUF Q4_K_M | 4.89 | 69.8 GB |
| GGUF Q3_K_M | 3.91 | 55.8 GB |
| GGUF Q2_K | 3.35 | 47.8 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× · 32.9 tok/s | 2× · 43.0 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 20.2 tok/s | 2× · 27.5 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 2× · 26.5 tok/s | 4× · 32.4 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 22.2 tok/s | 1× · 16.5 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 15.8 tok/s | 2× · 18.1 tok/s |
| L40S 48GBFP8 (W8A8) | 4× · 12.3 tok/s | 8× · 11.3 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 2× · 12.9 tok/s | 4× · 14.6 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 4× · 26.1 tok/s | 8× · 17.5 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 4× · 18.2 tok/s | 8× · 13.0 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 4.2 tok/s | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |