# MiniMax M3

A frontier-class agentic model that combines a long context window with native multimodality. MiniMax M3 accepts text, image, and video input and is tuned for multi-turn collaboration on coding and computer-use tasks.

- **Model ID:** `minimax/minimax-m3`
- **Type:** chat
- **Providers:** minimax, fireworks, nebius, gmicloud
- **Context window:** 1,049,000
- **Maximum output tokens:** 1,049,000
- **Pricing:** $0.24/1M input tokens, $0.96/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/minimax-m3

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

```ts
import { streamText } from 'ai'

const result = streamText({
  model: 'minimax/minimax-m3',
  prompt: 'Why is the sky blue?'
})
```

## About

MiniMax M3 is built around MiniMax Sparse Attention (MSA), an attention variant that splits the key-value cache into blocks and pre-filters which blocks contribute to each query. That design supports the 1.0M tokens context window without the quadratic compute scaling of full attention, and it lets MiniMax M3 keep prefill and decode efficient on long inputs.

Native multimodality is wired in from the start of training rather than bolted on later. MiniMax M3 accepts text, image, and video input and produces text output. The pretraining pipeline aligns visual and textual semantics directly, which carries over to multimodal coding tasks like analyzing a screenshot of a failing test and writing a patch, or reproducing a bug from a GitHub issue thread.

MiniMax M3 is positioned for software engineering, terminal-based tool use, and agentic web browsing. It scores 59.0% on SWE-Bench Pro and 70.06% on OSWorld-Verified for computer use. Automatic prompt caching is enabled by default, which reduces effective cost on repeated context patterns common in agent loops.

## What to consider

MiniMax M3 pairs the 1.0M tokens context window with native image and video input, which makes it a fit for workflows that reason over screenshots, design references, or long video transcripts alongside code. Route through AI Gateway via the AI SDK plus Chat Completions / Responses / Messages APIs to get provider failover, observability, and unified pricing across the providers serving MiniMax M3.

## When to use

### Best For

- Long-horizon coding agents that work across a full repository in a single session
- Multimodal engineering workflows that reason over screenshots, diagrams, or video walkthroughs alongside code
- Computer-use and browser agents that drive a desktop or web environment
- Terminal-driven tool chains where the agent reads command output and iterates over many steps
- Workloads that previously fragmented context across multiple requests because of window limits

### Consider Alternatives When

- Raw inference speed is more important than capability breadth (consider M3-highspeed)
- Your task is single-turn and text-only at short context length, where a smaller model is cheaper
- You only need text reasoning without multimodal input, where standard M2.7 covers the use case at lower cost

## Best for

- **Long-horizon coding agents:** Sessions that span an entire repository without fragmenting context across requests
- **Multimodal engineering:** Workflows that reason over screenshots, diagrams, or video alongside code
- **Computer-use agents:** Browser and desktop automation that benefits from strong OSWorld performance
- **Terminal-driven tool chains:** Agents that read command output and iterate across many steps
- **Long-video and long-document understanding:** Tasks that require sustained attention over hours of input

## Consider alternatives

- **Raw inference speed:** Latency matters more than capability breadth, so consider M3-highspeed
- **Short single-turn text tasks:** A smaller model is cheaper when the workload is single-turn and text-only
- **Text-only reasoning:** Standard M2.7 covers the use case at lower cost when multimodal input is not needed

## Frequently asked questions

### What is MiniMax Sparse Attention?

MSA is the attention variant behind MiniMax M3. It splits the key-value cache into blocks and pre-filters which blocks contribute to each query, which keeps compute manageable at the 1.0M tokens context length.

### What input types does MiniMax M3 accept?

Text, image, and video input. Output is text. Multimodality is native to MiniMax M3 rather than added through a separate vision adapter.

### What is the context window for MiniMax M3?

MiniMax M3 supports a context window of 1.0M tokens and a max output of 1.0M tokens per request.

### How does MiniMax M3 compare to M2.7?

M2.7 focuses on multi-agent orchestration, dynamic tool search, and text-only enterprise workflows. MiniMax M3 extends the series with native multimodal input, the 1.0M tokens context window, and the MSA architecture for long-context efficiency.

### Is there a faster variant of MiniMax M3?

Yes. Select `minimax/minimax-m3-highspeed` where your provider exposes it. The highspeed variant targets higher throughput with the same output behavior.

### Does MiniMax M3 support automatic prompt caching?

Yes. Automatic prompt caching is enabled by default, which reduces effective cost on repeated context patterns. $0.048 per million cached input tokens applies where the provider exposes a cached rate.

### How do I access MiniMax M3 through the AI SDK?

Set the model identifier to `minimax/minimax-m3` in your AI SDK configuration. AI Gateway routes the request across the providers serving MiniMax M3 with configurable failover.

### Is Zero Data Retention available for MiniMax M3?

Yes, Zero Data Retention is available for this model. Zero Data Retention is offered on a per-provider basis. See https://vercel.com/docs/ai-gateway/capabilities/zdr for details.

## Links

- [Model page](https://vercel.com/ai-gateway/models/minimax-m3)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://www.minimax.io/)