# MiMo M2.5

MiMo M2.5 is the standard tier of Xiaomi's MiMo v2.5 line. It's aimed at agentic workflows, code, and long-context analysis where multimodal input and tool use both matter.

- **Model ID:** `xiaomi/mimo-v2.5`
- **Type:** chat
- **Providers:** gmicloud
- **Context window:** 1,050,000
- **Maximum output tokens:** 131,100
- **Pricing:** $0.119/1M input tokens, $0.238/1M output tokens
- **Canonical page:** https://vercel.com/ai-gateway/models/mimo-v2.5

## Supported parameters

Detailed capability metadata has not been reported for this model.

## Example

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

const result = streamText({
  model: 'xiaomi/mimo-v2.5',
  prompt: 'Why is the sky blue?'
})
```

## About

MiMo M2.5 is a MoE language model from Xiaomi, released April 22, 2026 under the MIT license. Each forward pass activates a subset of total parameters, which keeps per-token compute lower than a dense model at the same parameter count.

The architecture uses hybrid attention, interleaving sliding-window and full attention to cut KV-cache storage at long sequence lengths. A multi-token prediction (MTP) block raises output tokens per step during inference. The full window of 1.1M tokens lets MiMo M2.5 reason over large documents, repos, or long agent trajectories.

MiMo M2.5 supports reasoning, tool calling, file input, vision, and implicit prompt caching. Call it through GMICloud via AI Gateway. For the higher-capability tier, see `mimo-v2.5-pro`.

## What to consider

MiMo M2.5 balances cost, capability, and context length. The MoE design keeps active compute small, but routing and serving a 300B-class MoE still requires capable infrastructure on the provider side. Use AI Gateway's cost tracking and model fallback to mix MiMo M2.5 with `mimo-v2.5-pro` on harder workloads.

## When to use

### Best For

- **Agentic workflows** that chain many tool calls in a single session
- **Software engineering** including refactors and repo-scale analysis within a window of 1.1M tokens
- **Multimodal input** mixing text, images, and uploaded files
- **Long-context analysis** over documents, codebases, or chat histories approaching 1.1M tokens
- **Cost-aware reasoning** where MoE active compute matters next to published scores

### Consider Alternatives When

- You need the deepest reasoning. `mimo-v2.5-pro` activates more parameters per step on harder problems
- You want throughput over reasoning depth. `mimo-v2-flash` is the speed-tuned sibling in the previous generation
- The task is simple classification or extraction where a smaller model is cheaper and fast enough
- Your inputs are always plain text and you don't need multimodal handling

## Best for

- **Agentic Workflows:** Tool-using agents that string together many calls in one session
- **Software Engineering:** Code generation, refactors, and repo-scale analysis with a window of 1.1M tokens
- **Multimodal Input:** Reasoning over mixed text, images, and uploaded files
- **Long Context:** Documents, codebases, or chat histories that approach 1.1M tokens
- **Cost-Aware Reasoning:** Lower active-parameter compute than dense models at similar scores

## Consider alternatives

- **Maximum Reasoning Depth:** `mimo-v2.5-pro` activates more parameters per step on the hardest math and engineering tasks
- **Speed-First Throughput:** `mimo-v2-flash` is the throughput-tuned option in the previous generation
- **Simple Classification:** A smaller, cheaper model handles short extraction at lower cost
- **Strict Text Pipelines:** A text-only model is fine when your inputs are never images or files

## Frequently asked questions

### How does MiMo M2.5 differ from `mimo-v2.5-pro`?

MiMo M2.5 is the standard tier. `mimo-v2.5-pro` activates a larger share of a larger parameter pool per step, so it costs more but reasons further on harder math, code, and agentic problems.

### What architecture does MiMo M2.5 use?

A Mixture-of-Experts (MoE) stack with hybrid attention. Each token activates a subset of expert blocks, and sliding-window plus full attention combine to keep KV cache manageable at long context lengths.

### What's the context window for MiMo M2.5?

1.1M tokens. Hybrid attention keeps KV-cache memory in check so long-context runs stay practical.

### Does MiMo M2.5 support tool calling and reasoning modes?

Yes. MiMo M2.5 supports tool calling and a reasoning mode, both exposed through AI Gateway. Use them through the AI SDK, the Chat Completions API, the Responses API, or any other supported format.

### How do I authenticate requests to MiMo M2.5 through AI Gateway?

Add your API key in AI Gateway project settings. Use `xiaomi/mimo-v2.5` in API calls. AI Gateway routes, retries, and fails over across `gmicloud`.

### What does MiMo M2.5 cost?

See the pricing section on this page for today's rates. AI Gateway tracks each provider's pricing for MiMo M2.5, so the numbers shown stay current.

### Does MiMo M2.5 support zero data retention?

Zero Data Retention is not currently 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.

### Is MiMo M2.5 available under an open-source license?

Yes. The MiMo v2.5 line is released under the MIT license, which allows commercial use, modification, and redistribution.

## Links

- [Model page](https://vercel.com/ai-gateway/models/mimo-v2.5)
- [AI Gateway documentation](https://vercel.com/docs/ai-gateway)
- [Provider model documentation](https://console.gmicloud.ai/user-console/ie/model-hub/llm/302d1538-75be-4c10-81a0-7e9037056e59)


If you encounter incorrect information, can’t find what you need, need a capability Vercel doesn’t currently offer, or have any other feedback to report, please do it by following the instructions at https://vercel.com/feedback.md.