MiniMax · released 2026-06 · 427B parameters (23.0B active per token)
MiniMax Sparse Attention on 57 of 60 layers (block-sparse: top-16 blocks of 128 chosen by a small 4-head x 128 indexer); full GQA KV (4 x 128) is still cached on every layer (indexer keys not counted). Card: ~428B total / ~23B active, 1M context; BF16 release (MXFP8 variant).
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.47 GB | 0.23 GB |
| 8,192 tokens | 0.94 GB | 0.47 GB |
| 32,768 tokens | 3.75 GB | 1.88 GB |
| 131,072 tokens | 15.0 GB | 7.50 GB |
| 1,048,576 tokens | 120.0 GB | 60.0 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 795.4 GB |
| FP8 (W8A8) | 8 | 402.3 GB |
| INT8 (W8A8 / W8A16) | 8 | 402.3 GB |
| NVFP4 | 4.5 | 230.3 GB |
| INT4 AWQ / GPTQ | 4.25 | 218.0 GB |
| GGUF Q8_0 | 8.5 | 422.6 GB |
| GGUF Q6_K | 6.56 | 326.1 GB |
| GGUF Q5_K_M | 5.69 | 282.9 GB |
| GGUF Q4_K_M | 4.89 | 243.1 GB |
| GGUF Q3_K_M | 3.91 | 194.4 GB |
| GGUF Q2_K | 3.35 | 166.5 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) | 4× · 149 tok/s | 4× · 45.3 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 4× · 129 tok/s | 4× · 30.0 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 8× · 84.5 tok/s | 8× · 31.9 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 4× · 133 tok/s | 4× · 32.5 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 4× · 117 tok/s | 4× · 23.3 tok/s |
| L40S 48GBFP8 (W8A8) | 16× · 20.5 tok/s | 16× · 10.6 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 8× · 38.8 tok/s | 8× · 14.6 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 16× · 20.9 tok/s | 16× · 15.3 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 16× · 20.4 tok/s | 16× · 12.4 tok/s |
| Apple M4 Max (40-core GPU, up to 128GB)GGUF Q4_K_M | > 16 | > 16 |
| DGX Spark (GB10, 128GB unified)FP8 (W8A8) | > 16 | > 16 |