MiniMax Open Platform

What is the MiniMax open platform?

The MiniMax open platform is a multi-modal model service platform designed for developers and enterprises, covering language, video, voice, images, and music.

Developers can integrate the MiniMax model into applications, agents, content creation processes, and programming workflows through APIs, SDKs, MCP, and the official CLI.

Key capabilities

• Invoke the long-context language model.
• Build code assistants and Agents.
• Use tools to make calls and perform searches on the server side.
• Generate video from text and video from images.
• Generation using initial and final frames along with multimodal references.
• Synchronous and asynchronous speech synthesis have been completed.
• Design or quickly replicate timbres.
• Generate images and stylized images.
• Manage file uploads and generation.
• It is invoked through an OpenAI-compatible interface.
• It is invoked through an Anthropic-compatible interface.
• Use MCP to connect multimodal tools.
• Use the CLI to generate content from the terminal.
• Deploy a partially open-weight model.

Current model framework

Which developers are it suitable for

MiniMax M3

MiniMax M3 is a cutting-edge language model designed for coding and Agent workflows, offering support for ultra-long contexts of up to 1M tokens as well as native multi-modal capabilities.

• Handle large code repositories and long documents.
• Execute multi-step Agent tasks.
• Perform tool calls and interleaved thought chains.
• Online searches are carried out using server-side tools.
• Use Prompt caching to reduce the cost of repetition.
• Compatible with OpenAI and Anthropic interface formats.

MiniMax M2.7 and the high-speed version

M2.7 and M2.7-highspeed offer the same level of performance; the high-speed version provides faster response times, but its unit cost is higher.

MiniMax H3 video model

MiniMax H3 supports multi-modal context understanding of text, images, videos, and audio, and generates videos with native stereo audio.

H3 Max is trained and has its speed optimized by a third party based on H3; its range of functions differs from that of the full H3 version, so it is necessary to check the model name before using it.

MiniMax Speech 2.8

Image model

Image-01 supports text-to-image and image-to-image generation; Image-01-live enhances styles such as hand-drawn and cartoon styles and allows for style settings.

• Create articles and social media images.
• Create visual variants based on the reference image.
• Create illustrations and content in a cartoon style.
• Create product concept diagrams quickly.
• Integrate MCP into the Agent workflow.

MiniMax API Integration Tutorial

1. Register and log in to the MiniMax open platform.
2. Complete the authentication required by the account.
3. Go to the console to create an API Key.
4. Choose between pay-as-you-go Key or subscription Key.
5. Do not expose keys in the browser frontend.
6. Read the latest documentation for the target model.
7. Choose OpenAI, Anthropic, or the official API.
8. Send the smallest possible request in the testing environment.
9. Record the model, tokens, and response time.
10. Add timeout, backoff, and retry mechanisms.
11. Poll the status of asynchronous video tasks.
12. Download the generated file in a timely manner.
13. Set balance warning thresholds and call limits.
14. Complete content security review before going live.
15. Review the bill based on the actual usage.

What are the differences between API keys?

The Token Plan does not cover a small number of special models such as H3, timbre design, and rapid replication; for these functionalities, regular pay-as-you-go keys should be used.

OpenAI-compatible interface

The open platform supports compatible interfaces such as OpenAI Chat Completions, Responses, and model lists, facilitating the migration of existing applications.

• Change the base address to the MiniMax endpoint.
• Replace with the MiniMax API Key.
• Use the platform’s current model name.
• Check for incompatible parameters and tool formats.
• Re-test the streaming output and error handling.
• Check the tokens and bills after migration.

Anthropic-compatible interface

Developers can also use the Anthropic SDK and the Messages format to invoke the MiniMax language model, which is suitable for tools within the Claude ecosystem.

• Configure compatible endpoints and authentication details.
• Confirm the system, messages, and tools formats.
• Check for support of interleaved thought chains.
• Avoid mixing different SDK fields simultaneously.
• Use caching to reduce the costs associated with long contexts.

Pay-as-you-go pricing for language models

Language models are charged separately for input, output, and caching, and M3 is also categorized based on whether a single input exceeds 512k tokens.

The current price page for M3 indicates a permanent 50% discount; however, the actual price should be based on the official pricing page as of the day of purchase.

Pay-as-you-go voice pricing

One Chinese character is counted as 2 characters, while English letters, spaces, punctuation marks, and line breaks are each counted as 1 character. There may also be costs associated with voice synthesis for audio playback.

Pay-as-you-go video pricing

Up to 5 images can be uploaded for free, as well as audio content via H3; any additional images or videos will be charged in accordance with the standard rates.

Prices for images and server-side tools

Token Plan subscription price

The Token Plan offers a monthly quota through subscription keys; text, images, and audio share the same quota, with time limits set at 5 hours and on a weekly basis.

The number of agents is only a reference for official scenarios; it does not guarantee a specific number of concurrent operations or requests. The actual usage depends on the context and the intensity of the tasks.

Scope of the Token Plan

Price of the points package

1,000 points are equivalent to approximately 7 yuan; the validity period is 365 days from the date of each purchase, with the package credits being used first before points are utilized.

Price of voice resource packages

The resource package also includes different RPM values as well as the number of free sound colors available; for large-scale operations, it is necessary to evaluate the validity period, concurrent usage, and consumption rate together.

Price of video resource packages

The current video resource package is available only for the Hailuo series and not for the MiniMax H3; for H3, pay-as-you-go APIs or custom business solutions should be used.

How to choose between pay-as-you-go, Token Plan, and resource packages

Current status of the music API

As of August 20, 2026, the paid music generation and lyric API will no longer be available to new users, and the free music generation service has also been discontinued.

Existing paid users can continue to use the current APIs; new users can use MiniMax Audio or MiniMax Music 3, for which the relevant weights have been made available.

MiniMax MCP

The official MCP tools offer multi-modal capabilities such as speech synthesis, voice cloning, and image and video generation, facilitating unified invocation by Agents.

• Configure the service in the compatible client.
• Use environment variables to save the API Key.
• Prepare the input file in accordance with the tool constraints.
• Add confirmation for paid generation operations.
• Record the task identifier and generation cost.
• Do not submit the key to the code repository.

MiniMax CLI

The official MiniMax CLI allows access to text, image, video, voice, and music capabilities from a terminal or an AI Agent.

1. Install MiniMax CLI through official channels.
2. Check the command version and operating environment.
3. Use interactive login or a secure configuration key.
4. First, check the current authentication status.
5. Test with low-cost text tasks.
6. Before generating the video, make sure to use a regular pay-as-you-go key.
7. Include paid operations in the budget and approval process.
8. Recheck the model parameters after upgrading the CLI.

Rate limiting and reliability

Different models and packages have their own separate RPM, TPM, and concurrency limits; once these limits are reached, it is necessary to wait and attempt the action again according to the backoff strategy.

• Do not immediately retry throttled requests indefinitely.
• Save the task_id for asynchronous tasks.
• Set business transaction serial numbers such as those for exponentiation.
• Distinguish between client errors and server errors.
• Set up a downgrade model for production operations.
• Monitor latency, failure rate, and costs.

Voice cloning security

To replicate a sound timbre, individual or corporate certification is required, as well as explicit authorization from the owner of that sound; it cannot be used for deception or fraud.

• Keep the proof of tone licensing.
• Do not create replicas of public figures without permission.
• It is prohibited from being used to bypass authentication.
• Synthesized speech should clearly indicate its AI origin.
• Establish processes for removing and disabling timbres.
• Manual review is added for high-risk scenarios.

Video and image compliance

Open weights and self-hosting

MiniMax has made the weights of certain language, video, and music models available, along with documentation for local or self-hosted deployment.

How to determine an open-source license

Different models, CLI tools, and example projects use various licenses; therefore, it is not possible to categorize all of them under the Apache, MIT, or unrestricted commercial license.

• Check each license in the model repository one by one.
• H3 uses a dedicated community license.
• Confirm commercial use, distribution, and usage restrictions.
• Model weights and platform APIs are determined separately.
• Third-party quantitative versions do not represent official authorization.
• The terms are reviewed by the legal department before the enterprise proceeds with the deployment.

Data and key security

• The API Key is stored only on the server side.
• Different keys are used for development, testing, and production.
• Set permissions and define key rotation procedures.
• Remove sensitive information before uploading the file.
• Delete files that are no longer needed in a timely manner.
• Do not record the full key in the logs.
• Set balance alerts for abnormal calls.

Product advantages

• Five types of multimodal models are provided uniformly.
• It supports interfaces compatible with OpenAI and Anthropic.
• M3 is suitable for long-context programming and Agents.
• H3 supports the creation of complex, multimodal videos.
• Voice services offer synthesis, design, and replication.
• Options for pay-as-you-go, subscription, and resource packs are all available.
• MCP, CLI, and OpenAPI documentation are provided.
• Some models can be self-hosted locally.

Usage restrictions

• Different Keys cover different resources.
• The Token Plan does not include special models such as H3.
• There may be an additional fee for video input materials.
• Resource packages have a validity period and RPM limits.
• The music API is no longer available to new users.
• Open models require high computing power.
• Permits for different warehouses are not the same.
• The generated content still requires manual review.

Frequently Asked Questions

What models does the MiniMax open platform offer?

The platform supports various language and multimodal models such as MiniMax M3, M2.7, H3, Speech 2.8, image-01, and Music 3.

Does the MiniMax API support the OpenAI format?

Yes, the open platform offers interfaces compatible with OpenAI Chat Completions and Responses, as well as support for the Anthropic Messages format.

What is the pay-as-you-go price for the MiniMax M3?

For standard services, when the input length is within 512k characters, the cost is 2.1 yuan for the input and 8.4 yuan per million tokens for the output; higher prices apply for longer inputs.

How much is the MiniMax Token Plan?

Plus, Max, and Ultra cost 49 yuan, 119 yuan, and 469 yuan per month respectively, but features such as H3 and tone replication are not included in these packages.

How much is the MiniMax H3 video?

At present, for H3, the cost is 0.50 yuan per second for 768P resolution and 0.80 yuan per second for 2K resolution; uploading videos or more images than the free quota allows may incur additional charges.

Is it still possible to apply for the MiniMax music API?

New users are not allowed to use them; as of August 20, 2026, the paid music and lyrics APIs will no longer be available to new users, and the free interfaces have also been discontinued.

Is MiniMax an open-source platform?

The open platform represents commercial API services; however, certain models such as M3, M2.7, H3, and Music 3 offer open weights, and the licenses for these need to be checked individually.