MiniMax-M3 is a native multimodal model with 1M context. It has ~428B parameters and ~23B activated parameters.
Highlights:
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.
model-00059-of-00059.safetensors 16 GBmodel-00010-of-00059.safetensors 16 GBmodel-00011-of-00059.safetensors 16 GBmodel-00012-of-00059.safetensors 16 GBmodel-00013-of-00059.safetensors 16 GBmodel-00014-of-00059.safetensors 16 GBmodel-00015-of-00059.safetensors 16 GBmodel-00016-of-00059.safetensors 16 GBmodel-00017-of-00059.safetensors 16 GBmodel-00018-of-00059.safetensors 16 GBmodel-00019-of-00059.safetensors 16 GBmodel-00020-of-00059.safetensors 16 GBmodel-00021-of-00059.safetensors 16 GBmodel-00022-of-00059.safetensors 16 GBmodel-00023-of-00059.safetensors 16 GBmodel-00024-of-00059.safetensors 16 GBmodel-00003-of-00059.safetensors 16 GBmodel-00004-of-00059.safetensors 16 GBmodel-00005-of-00059.safetensors 16 GBmodel-00006-of-00059.safetensors 16 GBmodel-00007-of-00059.safetensors 16 GBmodel-00008-of-00059.safetensors 16 GBmodel-00009-of-00059.safetensors 16 GBmodel-00026-of-00059.safetensors 15 GBmodel-00027-of-00059.safetensors 15 GBmodel-00028-of-00059.safetensors 15 GBmodel-00029-of-00059.safetensors 15 GBmodel-00030-of-00059.safetensors 15 GBmodel-00031-of-00059.safetensors 15 GBmodel-00032-of-00059.safetensors 15 GBmodel-00033-of-00059.safetensors 15 GBmodel-00034-of-00059.safetensors 15 GBmodel-00035-of-00059.safetensors 15 GBmodel-00036-of-00059.safetensors 15 GBmodel-00037-of-00059.safetensors 14 GBmodel-00038-of-00059.safetensors 14 GBmodel-00039-of-00059.safetensors 14 GBmodel-00040-of-00059.safetensors 14 GBmodel-00041-of-00059.safetensors 14 GBmodel-00042-of-00059.safetensors 14 GBmodel-00043-of-00059.safetensors 14 GBmodel-00044-of-00059.safetensors 14 GBmodel-00045-of-00059.safetensors 14 GBmodel-00046-of-00059.safetensors 14 GBmodel-00047-of-00059.safetensors 14 GBmodel-00048-of-00059.safetensors 14 GBmodel-00049-of-00059.safetensors 14 GBmodel-00050-of-00059.safetensors 14 GBmodel-00051-of-00059.safetensors 14 GBmodel-00052-of-00059.safetensors 14 GBmodel-00053-of-00059.safetensors 14 GBmodel-00054-of-00059.safetensors 14 GBmodel-00055-of-00059.safetensors 14 GBmodel-00056-of-00059.safetensors 14 GBmodel-00057-of-00059.safetensors 14 GBmodel-00058-of-00059.safetensors 14 GBmodel-00002-of-00059.safetensors 11 GBmodel-00001-of-00059.safetensors 5.6 GBmodel-00025-of-00059.safetensors 5.2 GB有文件？提交种子。
提交种子
LocalMaxxingmlx_lm.server --model MiniMax-M3-MLX-3bit --port 8096 --chat-template-args {thinking_mode: disabled}在 LocalMaxxing 上查看测试 docker run -d --name lsv-mmm3-iq1m-v16 --gpus all --ipc=host -p 127.0.0.1:8000:8000 -v /mnt/models:/mnt/models -v /mnt/models/cache:/root/.cache -e HF_HOME=/mnt/models/hf ghcr.io/ggml-org/llama.cpp@sha256:b51c252687249a6990ae1c168156cb215c954bfd71f9c434eda9a065b1a40dbe -m /mnt/models/gguf/MiniMax-M3-IQ1M/UD-IQ1_M/MiniMax-M3-UD-IQ1_M-00001-of-00004.gguf --host 0.0.0.0 --port 8000 -fa on -c 32768 -np 8 --jinja -t 8 -tb 16 --metrics -b 4096 -ub 2048 -lm none在 LocalMaxxing 上查看测试
来自 LocalMaxxing 的 2 次社区测试（2 种配置）的中位数。
自己运行：配置指南