MiniMax M3
Neural Network
MiniMax M3 is a multimodal model with a 1M token context: MiniMax M3 API for agentic scenarios and coding tasks.
Specifications
- Input
- text, images, documents
- Output
- text
- Context size
- 1 048 576 tokens
- Max answer length
- 512 000 tokens
- Reasoning effort
- low, medium, high
- Capabilities
- reasoning, function calling
Prompt cost
(per 1M tokens)
Answer cost
(per 1M tokens)
How it works MiniMax M3?
MiniMax M3 is a multimodal base model by MiniMax: it accepts text, images, and video as input and responds with text. The 1M token context allows you to keep a large codebase, a long document, or an entire agent history in a single dialogue, making the model ideal for long agentic scenarios and programming. Through BotHub, the model is accessible from Russia without a VPN or foreign card, and you pay in rubles as you go—for tokens used, without a subscription. With over 250 neural networks in one window, it's easy to compare responses for a single prompt, and a unified OpenAI-compatible API allows you to switch models without rewriting your integration. Chats are encrypted via AES-GCM; for companies, we offer contracts, invoices, EDI, and an admin panel with limits. It is useful when you need to analyze a large repository and refactor code, build an agent for a multi-step task without losing context, extract meaning from screenshots and interface videos, or prepare technical documentation from various formats.Frequently asked questions about MiniMax M3
You can use generated results for commercial purposes. You own all rights to the content you create. The only restriction: make sure your prompt does not include copyrighted third-party material. You are responsible for respecting the rights to any input data.
It is a text model with a reasoning mode: parsing large documents and code, analytics, STEM tasks, refactoring, and debugging. The context window is 1,048,576 tokens, so a large codebase or a stack of reports will fit into a single dialogue. It accepts text, documents, and images as input.
The single response limit is 512,000 tokens, which is a very large volume. It is enough for long documentation, detailed analysis, or a large code module in its entirety, without splitting it into parts. In practice, the length depends on the prompt: explicitly request the desired format and volume, otherwise the model will respond briefly.
Yes, reasoning before answering is a stated capability of the model. It first builds a chain of steps and only then provides the result—this significantly helps in mathematics, logic, requirements analysis, and debugging, where it is important not to guess, but to consistently break down the task.
Yes, the model supports function calling and structured JSON output. It is convenient to use as the basis for agents and pipelines: you describe tools and receive a predictable response according to the schema. In BotHub, this works via a unified OpenAI-compatible API, so switching to another model does not break the integration.
Images and documents—yes, they can be sent along with text: interface screenshots, diagrams, tables, PDF reports. Audio and video are not marked as input formats in our data, so for working with sound or video recordings, it is better to choose a specialized model from the catalog.
There is no separate assessment of Russian language quality in our data, so we will not promise specific figures. The easiest way to check is with your own task: send a typical request and compare the result with another model in the same window—access from Russia is available without a VPN or foreign card.