Mistral Small · released 2026-03 · 119B parameters (6.50B active per token)
MLA (KV latent 256 + 64 rope = 320 elements/token/layer) + 128-expert MoE; card: 119B total, 6.5B active, 256K context (config max_position_embeddings 1048576). Includes vision encoder. NVFP4 variant published.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.09 GB | 0.04 GB |
| 8,192 tokens | 0.18 GB | 0.09 GB |
| 32,768 tokens | 0.70 GB | 0.35 GB |
| 131,072 tokens | 2.81 GB | 1.41 GB |
| 262,144 tokens | 5.63 GB | 2.81 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 222.4 GB |
| FP8 (W8A8) | 8 | 113.2 GB |
| INT8 (W8A8 / W8A16) | 8 | 113.2 GB |
| NVFP4 | 4.5 | 65.4 GB |
| INT4 AWQ / GPTQ | 4.25 | 62.0 GB |
| GGUF Q8_0 | 8.5 | 118.2 GB |
| GGUF Q6_K | 6.56 | 91.2 GB |
| GGUF Q5_K_M | 5.69 | 79.1 GB |
| GGUF Q4_K_M | 4.89 | 68.0 GB |
| GGUF Q3_K_M | 3.91 | 54.3 GB |
| GGUF Q2_K | 3.35 | 46.6 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× · 346 tok/s | 1× · 50.4 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 251 tok/s | 1× · 31.0 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 2× · 253 tok/s | 2× · 40.0 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 268 tok/s | 1× · 34.1 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 207 tok/s | 1× · 23.6 tok/s |
| L40S 48GBFP8 (W8A8) | 4× · 95.8 tok/s | 4× · 18.0 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 2× · 139 tok/s | 2× · 19.3 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 4× · 110 tok/s | 4× · 39.6 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 4× · 98.5 tok/s | 8× · 29.0 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 62.3 tok/s | 1× · 6.4 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |