Can you run MiniMax M2.5 locally?
Yes. MiniMax’s M2.5 repository documents a vLLM path for serving MiniMaxAI/MiniMax-M2.5 on your own infrastructure. That is local self-hosting—not the same as calling a hosted API.
The vLLM guide lists Linux, Python 3.9–3.12, a GPU with compute capability 7.0 or higher, 220 GB for model weights, and 240 GB per 1 million context tokens. It also provides the four-GPU serving example below. Those details do not establish that any four-GPU machine will have sufficient memory; check the guide against your hardware and workload before provisioning.
vLLM serving command
This is the model-specific serving command from MiniMax’s guide. It is not a complete installation walkthrough. Follow the official guide for environment setup and its health-check request.
SAFETENSORS_FAST_GPU=1 vllm serve \
MiniMaxAI/MiniMax-M2.5 --trust-remote-code \
--tensor-parallel-size 4 \
--enable-auto-tool-choice --tool-call-parser minimax_m2 \
--reasoning-parser minimax_m2_append_think
For production, confirm the model revision and runtime you intend to use, review the model license, and secure the service before exposing it beyond your machine. With self-hosting, you provide and operate the infrastructure and runtime, and manage capacity, scaling, monitoring, recovery, and upgrades.
Local inference versus API access
Local inference runs the model and inference service on infrastructure you control. An API request goes to a hosted service; it does not mean the model weights are running on your hardware. MiniMax’s general deployment guidance recommends its API for a managed service or elastic scaling, but that guidance alone does not confirm that M2.5 is currently available through the API. Verify the M2.5 model identifier and access details in current official API documentation before choosing that route.
Do not confuse M2.5 with MiniMax M2 or M2.7: their commands and deployment requirements are not M2.5 instructions.