Ministral 3 · released 2025-12 · 8.92B parameters
Vision-language (incl. ~0.4B vision encoder). Main Instruct repo ships FP8 weights; params from the BF16 repo.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.53 GB | 0.27 GB |
| 8,192 tokens | 1.06 GB | 0.53 GB |
| 32,768 tokens | 4.25 GB | 2.13 GB |
| 131,072 tokens | 17.0 GB | 8.50 GB |
| 262,144 tokens | 34.0 GB | 17.0 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 16.6 GB |
| FP8 (W8A8) | 8 | 10.3 GB |
| INT8 (W8A8 / W8A16) | 8 | 10.3 GB |
| NVFP4 | 4.5 | 7.55 GB |
| INT4 AWQ / GPTQ | 4.25 | 7.35 GB |
| GGUF Q8_0 | 8.5 | 8.82 GB |
| GGUF Q6_K | 6.56 | 6.81 GB |
| GGUF Q5_K_M | 5.69 | 5.91 GB |
| GGUF Q4_K_M | 4.89 | 5.08 GB |
| GGUF Q3_K_M | 3.91 | 4.06 GB |
| GGUF Q2_K | 3.35 | 3.48 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× · 260 tok/s | 1× · 127 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 178 tok/s | 1× · 81.2 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 133 tok/s | 1× · 58.4 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 192 tok/s | 1× · 88.8 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 114 tok/s | 1× · 40.8 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 38.9 tok/s | 1× · 15.9 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 69.2 tok/s | 1× · 28.9 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 108 tok/s | 1× · 33.1 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 68.8 tok/s | 2× · 33.7 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 47.6 tok/s | 1× · 10.0 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 12.7 tok/s | 1× · 5.1 tok/s |