Made it in MiniMax? Keep what's next moving.

Who makes MiniMax

MiniMax is an AI company based in Shanghai, China. Alongside large language models (LLMs) for text, it builds its own video generation and speech synthesis models and apps. Its international apps are provided by Nanonoble Pte. Ltd., a company registered in Singapore.

Three ways to use it

• Apps: MiniMax Code, MiniMax Audio, Hailuo for video and more, on the web or your phone.
• API: build it into your own systems or developer tools through MiniMax's platform (platform.minimax.io). Pay-as-you-go and a flat-rate Token Plan are available.
• Open weights: run the MiniMax-M3 or MiniMax H3 weights published on Hugging Face on your own servers, subject to license terms.

"Minimax" is also the name of a search algorithm in game theory. This page is about MiniMax the AI company.

Scroll the table sideways →

What is MiniMax H3?

MiniMax H3 is a video generation model MiniMax announced in late July 2026. It takes text, images, video and audio as cues and produces video with stereo sound (4 to 15 seconds, up to 2K). The official model card says it supports 11 languages for spoken dialogue, including Japanese.

• Where to use it: the Hailuo video app, MiniMax's API, and the weights published on Hugging Face.
• How it differs from M3: H3 makes video; M3 is a text model. They are different models.
• Running it locally: it's heavy for a personal computer. See the reality of running it locally below.

What is MiniMax-M3?

MiniMax-M3 is MiniMax's current flagship text model, released in June 2026. It has about 428 billion parameters in total (about 23 billion active per inference), a 1M-token context, and accepts text, image and video input. API prices are low, and there's a Token Plan for flat-rate use from coding tools.

How to use MiniMax Audio

Paste the text you want read aloud, choose a voice and speaking speed, and generate the audio. It works for training video narration, product voiceovers and internal announcements. If you make a voice that resembles a real person, get their consent and check the rules of wherever you'll publish it.

• Text: draft scripts, articles, emails and reports with MiniMax-M3. It can also summarize long documents and explain images or video you show it.
• Voice: turn scripts into narration and voiceovers with MiniMax Audio, for training, product introductions and internal announcements.
• Video: use MiniMax H3 in Hailuo or via the API to make short videos with sound from text or images. Good for storyboard-style drafts and test clips for social media.
• Coding: hand off coding work to MiniMax Code, or to coding tools connected through the Anthropic-compatible API.

Making a video in Hailuo (H3)

1. Describe the scene in words. Be specific about place, people and action, such as "a morning office where staff hold a meeting while looking at a tablet".
2. Add an image if needed. You can animate a product photo or a reference image.
3. Make several and choose. The same prompt gives different results each time, so comparing a few candidates is the quickest path.

If you use faces, logos or anything resembling someone else's work, check the rights and the rules of where you'll publish. Have a person review any video before it goes outside the company.

Whatever the use, MiniMax delivers assets and drafts. The owners and deadlines decided in meetings, and the messages to the people involved, are what SIMY picks up from meetings and chat and moves forward.

• Pay-as-you-go: as of October 2026, MiniMax-M3 costs $0.30 per million input tokens and $1.20 per million output tokens for inputs up to 512K tokens. Longer inputs cost twice as much, and priority processing costs 1.5 times as much. That's on the cheap side even among Chinese LLMs.
• Token Plan: a flat-rate plan for use from coding tools and similar, in three tiers: Plus at $22 a month, Max at $55 and Ultra at $132. Each has usage limits per five hours and per week.
• Taxes: prices are in US dollars. Check the pay-as-you-go pricing page and the Token Plan page for tax treatment and the latest figures.

Which to choose

• You use it daily from coding tools: if your usage is hard to predict, the Token Plan makes the monthly bill easier to plan.
• You're building it into your own systems: pay-as-you-go, which scales with volume, fits better. Start small, check the cost per task, then expand.

Video and audio are priced separately from text and vary with the number and length of what you make, so check the official pages.

• Getting in: create an API key on MiniMax's platform (platform.minimax.io).
• Compatibility: MiniMax recommends a format compatible with Anthropic's SDK. It also supports OpenAI's Chat Completions and Responses API formats.
• From coding tools: point Claude Code or another Anthropic-format tool at MiniMax's endpoint and API key, and you can call MiniMax-M3. The official docs walk through it.

Getting started with the API

1. Create an account on MiniMax's platform. Choose pay-as-you-go or the Token Plan.
2. Create an API key. Keep it in an environment variable, never directly in code or shared documents.
3. Point your usual tools at it. Pick the Anthropic or OpenAI format, then swap in the URL and model name.
4. Start small and watch usage. Long inputs cost more per token. Check daily usage before scaling up to production volume.

How SIMY relates: SIMY doesn't connect MiniMax API keys at this time. SIMY's cloud runs use your ChatGPT (Codex) account.

The MiniMax-M3 weights are published on Hugging Face, but under MiniMax's own "MiniMax Community License". Non-commercial use is free; commercial use comes with conditions.

• Commercial use: you must display "Built with MiniMax M3" and notify MiniMax.
• Annual revenue above $20 million: you need MiniMax's written permission before commercial use.
• Prohibited uses: military use and others are prohibited.
• M2.7: also under its own license.

The MiniMax H3 license

• Territory: the MiniMax H3 Community License excludes the EU, the UK, South Korea and the US from its territory. If you operate in or work with partners in those regions, check carefully before using it.
• Commercial use: you must show "MiniMax H3" in your product's interface. Companies with annual revenue above $20 million need written permission first.

Even for internal trials, the terms depend on whether the use counts as commercial. Before using it for work, read the M3 license text and the H3 license text.

Lots of people search for "minimax h3 local", but as of October 2026, expect it to be hard to run H3 or M3 on an ordinary computer.

• Ollama: minimax-m3 and minimax-m2.7 in the official library carry the "cloud" tag. That means they run in Ollama's cloud, and your input is sent to Ollama's servers. Community-built versions exist, but they aren't official.
• MiniMax H3: the weights are public, but the official instructions use a four-GPU setup as the example. There's also a guide for using it in ComfyUI; check the GPU memory it needs on your own hardware.
• Official guides for M3 and M2.7: M3 is shown on eight B200 GPUs (marked experimental), and M2.7 on four high-memory GPUs. These are data-center setups.
• If you want something on your own computer: the small text models covered in the Qwen guide and the GLM guide are options.

If the goal is keeping data in-house, cloud-tagged models don't qualify. Either run the weights on your own GPU servers or choose a model small enough to run locally.

Scroll the table sideways →

• Check your company's rules first: some companies' generative AI policies limit which services you can use and what you can enter.
• Keep confidential data out by default: with any AI, think twice before entering customer information or unpublished figures.
• Take care with voices and faces: when you use a real person's voice or face to generate audio or video, get their consent and check the rules of where you'll publish.

Data handling is based on the privacy policy for the international apps and the MiniMax privacy policy as of October 1, 2026. How SIMY handles data is published on our Security page and in our Privacy Policy.

You can prompt MiniMax-M3 in English and many other languages, but as of October 2026, M3's supported languages aren't officially listed. The H3 video model's card, on the other hand, lists 11 languages for spoken dialogue. Either way, test with real material from your own work.

• Text: check the tone and phrasing of scripts and email drafts before using them.
• Voice: listen for how names, product terms and numbers are read aloud.
• Video: check that any on-screen text isn't garbled. Adding subtitles afterwards in an editor is the safe route.

You can also chain it: translate a script into another language with M3, then voice it with MiniMax Audio. The owners and deadlines for the fixes agreed in the meeting, and the messages to the production company and your team, can be left to SIMY.

Scroll the table sideways →

A quick way to choose

• Text, voice and video in one place: MiniMax (H3 for video, M3 for text).
• Feeding it long documents in one go: see the Kimi guide.
• Lowest API cost: MiniMax or DeepSeek (guide). DeepSeek's prices vary by time of day.

How they fit together: any AI, MiniMax included, can make the assets and drafts. If you're chasing the owners, deadlines, requests and updates yourself every time after that, that's where SIMY comes in. See Chinese LLMs compared for the differences between providers, the Doubao guide for ByteDance's models, and the GLM guide for Zhipu's.