MiniMax · released 2025-06 · 456B parameters (45.9B active per token)
70 lightning (linear) attention layers + 10 softmax GQA layers (every 8th). 456B / 45.9B active; 1M input context, 80K thinking budget. Lightning state (64 heads x 128 x 128 per layer) assumed BF16.
Source configMemory for one conversation's context, on top of the weights.
| Context | FP16 | FP8 |
|---|---|---|
| 4,096 tokens | 0.29 GB | 0.21 GB |
| 8,192 tokens | 0.45 GB | 0.29 GB |
| 32,768 tokens | 1.39 GB | 0.76 GB |
| 131,072 tokens | 5.14 GB | 2.64 GB |
| 1,000,000 tokens | 38.3 GB | 19.2 GB |
| Format | Bits | Weights |
|---|---|---|
| BF16 / FP16 | 16 | 849.5 GB |
| FP8 (W8A8) | 8 | 429.3 GB |
| INT8 (W8A8 / W8A16) | 8 | 429.3 GB |
| NVFP4 | 4.5 | 245.5 GB |
| INT4 AWQ / GPTQ | 4.25 | 232.4 GB |
| GGUF Q8_0 | 8.5 | 451.3 GB |
| GGUF Q6_K | 6.56 | 348.3 GB |
| GGUF Q5_K_M | 5.69 | 302.1 GB |
| GGUF Q4_K_M | 4.89 | 259.6 GB |
| GGUF Q3_K_M | 3.91 | 207.6 GB |
| GGUF Q2_K | 3.35 | 177.9 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× · 107 tok/s | 4× · 33.0 tok/s |
| H200 SXM 141GBFP8 (W8A8) | 4× · 89.1 tok/s | 4× · 21.7 tok/s |
| H100 SXM 80GBFP8 (W8A8) | 8× · 61.3 tok/s | 8× · 24.2 tok/s |
| Instinct MI300X 192GBFP8 (W8A8) | 4× · 92.7 tok/s | 4× · 23.6 tok/s |
| A100 SXM 80GBINT4 AWQ / GPTQ | 4× · 79.7 tok/s | 4× · 17.2 tok/s |
| L40S 48GBFP8 (W8A8) | 16× · 15.1 tok/s | 16× · 8.3 tok/s |
| RTX PRO 6000 Blackwell Server Edition 96GBFP8 (W8A8) | 8× · 28.0 tok/s | 8× · 11.0 tok/s |
| GeForce RTX 5090 32GBGGUF Q4_K_M | 16× · 15.7 tok/s | 16× · 12.1 tok/s |
| GeForce RTX 4090 24GBGGUF Q4_K_M | 16× · 15.3 tok/s | 16× · 10.0 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 |