Model reference · open weights
MiniMax-M3 is an open-weight language model from MiniMax. MiniMax-M3-MXFP8 (FP8) weighs 444 GB; the smallest configuration that runs it is 4× H200 141 GB.
Summary of the MiniMaxAI/MiniMax-M3-MXFP8 model card, 2026-10-01
What it is
| Released by | MiniMax |
|---|---|
| Released | 2026-06-02 |
| Parameters | 440.3B |
| VRAM | 444 GB for the weights |
What it runs on
| Card | Requests at once | Context max | Memory | |
|---|---|---|---|---|
| 8K each | 32K each | |||
| RTX 3060 12 GB … B200 180 GB 12 smaller cards | — | — | — | |
| 4× H200 141 GB tensor parallel | 80 | 20 | 643K | 138 GB a card |
| 8× H100 80 GB tensor parallel | 62 | 15 | 503K | 78.1 GB a card |
| 8× A100 80 GB tensor parallel · FP8 without its speed-up here | 82 | 20 | 657K | 78.2 GB a card |
| 4× B200 180 GB tensor parallel | 230 | 57 | all 1024K | 176 GB a card |
| Requests at once | 8K tokens each | 32K tokens each |
|---|---|---|
| 1 | 447 GB | 450 GB |
| 5 | 451 GB | 466 GB |
| 8 | 454 GB | 478 GB |
| 16 | 462 GB | 510 GB |
| 32 | 478 GB | 575 GB |
| 64 | 510 GB | 703 GB |
One card, with vLLM's small-card settings.
Builds
| Build | Params | Precision | Weights | Smallest setup |
|---|---|---|---|---|
| MiniMax-M3-MXFP8 (above) ↗ | 440.3B | FP8 | 444 GB | 4× H200 141 GB |
| MiniMax-M3 ↗ | 427.0B | BF16 | 854 GB | 8× H200 141 GB |
| MiniMax-M3-NVFP4 ↗ packaged by nvidia |
220.5B | NVFP4 | 250 GB | 2× H200 141 GB |
| MiniMax-M3-GGUF ↗ packaged by unsloth 22 builds: IQ1_M 128 GB … BF16 852 GB |
— | GGUF | 264 GB | 2× B200 180 GB |
From the model card
MiniMax-M3 is a native multimodal model with 1M context. It has ~428B parameters and ~23B activated parameters.
Highlights:
MiniMax-M3-MXFP8 is the MXFP8 quantized variant of MiniMax-M3, a native multimodal model with 1M context. It has ~428B parameters and ~23B activated parameters.
M3 is powered by MiniMax Sparse Attention (MSA), a high-performance sparse attention operator designed for million-token contexts. Compared with GQA, MSA dramatically reduces the attention compute and memory footprint while preserving model quality.
📄 Read the technical report: arXiv:2606.13392 · Hugging Face Papers
M3 supports three reasoning modes through the thinking parameter:
enabled — Reasoning is always enabled.adaptive — M3 automatically determines when additional reasoning is beneficial.disabled — Reasoning is disabled to minimize latency and maximize throughput.Download the model:
hf download MiniMaxAI/MiniMax-M3 --local-dir MiniMax-M3
We recommend the following inference frameworks to serve the model:
SGLang - see SGLang cookbook.
vLLM - see vLLM recipes.
Transformers - see Transformers docs.
We recommend the following parameters for best performance: temperature=1.0, top_p=0.95.
Contact us at model@minimax.io.
Quoted from the model card on Hugging Face. The full card is behind the Hugging Face link above.