Skip to main content
MiniMax provides open-weight models for language, multimodal understanding, music generation, and audiovisual generation. This guide summarizes the official deployment baselines and links to an end-to-end workflow for each model.
This section primarily covers self-hosted deployment on servers or clusters. The H3 page also provides a ComfyUI local-workflow path. You provide the hardware, operate the runtime, configure access control when applicable, and implement content safeguards. For a managed service, elastic scaling, or the complete MiniMax Platform feature set, use the MiniMax API.

​
Choose a model

The status describes the maturity of the documentation baseline, not the product lifecycle of the model itself:
  • Stable: Uses a fixed runtime version and provides a reproducible reference command.
  • Preview: An official deployment path is available, but the upstream installation method or some capabilities are still changing.
  • Experimental: Depends on a prerelease runtime, or some hardware combinations have not completed end-to-end validation.
“Open weights” does not mean that every model uses the same license. Before deployment, redistribution, or commercial use, read the complete License and Acceptable Use Policy linked from the model page. Inference-framework support does not grant additional model usage rights.

​
Self-hosting boundaries

Self-hosting keeps model weights, input data, and the inference service on infrastructure that you control. It also means that:
  • You are responsible for capacity planning, scaling, monitoring, recovery, and upgrades.
  • Quickstarts listen only on the local machine by default. Before exposing a service, configure authentication, TLS, network isolation, and remote-media access restrictions.
  • Open weights do not include MiniMax Platform managed files, caching, content safeguards, or platform-only workflows.
  • Community quantizations, converted weights, and other inference frameworks are outside this section’s validated scope unless a model page explicitly lists them.

​
Before you deploy

  1. Read the target model’s License, open-release scope, and verified capabilities.
  2. Prepare the GPU, host memory, disk, driver, and CUDA environment specified on the model page.
  3. Pin the model revision, inference runtime version, or image digest. Do not let production deployments drift with main, dev, or latest.
  4. Prepare a Hugging Face cache or offline weights. For mirrors in mainland China, use only the official source listed on the model page.
  5. Complete the health check and minimal request before applying quantization, parallelism, offloading, or throughput tuning.
Each model page provides the reference baseline adopted by the MiniMax documentation. SGLang Cookbook configurators can generate additional hardware, quantization, and topology combinations, but a generated result may be marked Unverified. Do not reuse parameters from different hardware without regenerating the complete configuration.

​
Support and validation scope

Each model page identifies the model revision, runtime version, reference hardware, verified capabilities, and last validation date. When official measurements are unavailable, the page says “not published” or “unverified” rather than estimating from adjacent models or community results.