Mistral Medium · released 2026-04 · 128B parameters
Dense 128B with a from-scratch vision encoder (~2.7B by difference vs Devstral 2's identical text stack). Ships FP8 weights.
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 |
| 262,144 tokens | 88.0 GB | 44.0 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 237.9 GB |
| FP8 (W8A8) | 8 | 124.9 GB |
| INT8 (W8A8 / W8A16) | 8 | 124.9 GB |
| NVFP4 | 4.5 | 75.5 GB |
| INT4 AWQ / GPTQ | 4.25 | 72.0 GB |
| GGUF Q8_0 | 8.5 | 126.4 GB |
| GGUF Q6_K | 6.56 | 97.5 GB |
| GGUF Q5_K_M | 5.69 | 84.6 GB |
| GGUF Q4_K_M | 4.89 | 72.7 GB |
| GGUF Q3_K_M | 3.91 | 58.1 GB |
| GGUF Q2_K | 3.35 | 49.8 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× · 30.6 tok/s | 2× · 41.0 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 2× · 34.1 tok/s | 2× · 26.1 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 2× · 24.7 tok/s | 4× · 31.0 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 20.7 tok/s | 2× · 28.5 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 2× · 25.8 tok/s | 2× · 16.8 tok/s |
| L40S 48GBFP8 (W8A8) | 4× · 11.6 tok/s | 8× · 11.0 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 2× · 12.0 tok/s | 4× · 14.0 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 4× · 25.5 tok/s | 8× · 17.3 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 4× · 17.7 tok/s | 8× · 12.9 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 4.0 tok/s | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |