Ministral 3 · released 2025-12 · 3.85B parameters
Vision-language (incl. ~0.4B vision encoder). Main Instruct repo ships FP8 weights; params from the BF16 repo. Vision encoder ~0.42B (~11% of total).
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.41 GB | 0.20 GB |
| 8,192 tokens | 0.81 GB | 0.41 GB |
| 32,768 tokens | 3.25 GB | 1.63 GB |
| 131,072 tokens | 13.0 GB | 6.50 GB |
| 262,144 tokens | 26.0 GB | 13.0 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 7.17 GB |
| FP8 (W8A8) | 8 | 3.96 GB |
| INT8 (W8A8 / W8A16) | 8 | 3.96 GB |
| NVFP4 | 4.5 | 2.56 GB |
| INT4 AWQ / GPTQ | 4.25 | 2.46 GB |
| GGUF Q8_0 | 8.5 | 3.81 GB |
| GGUF Q6_K | 6.56 | 2.94 GB |
| GGUF Q5_K_M | 5.69 | 2.55 GB |
| GGUF Q4_K_M | 4.89 | 2.19 GB |
| GGUF Q3_K_M | 3.91 | 1.75 GB |
| GGUF Q2_K | 3.35 | 1.50 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× · 459 tok/s | 1× · 190 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 1× · 346 tok/s | 1× · 125 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 1× · 273 tok/s | 1× · 90.9 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 1× · 368 tok/s | 1× · 136 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 1× · 260 tok/s | 1× · 62.9 tok/s |
| L40S 48GBFP8 (W8A8) | 1× · 91.1 tok/s | 1× · 25.4 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 1× · 155 tok/s | 1× · 45.8 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 1× · 184 tok/s | 1× · 46.6 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 1× · 127 tok/s | 1× · 27.5 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | 1× · 91.9 tok/s | 1× · 18.3 tok/s |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | 1× · 30.9 tok/s | 1× · 8.2 tok/s |