# MiniMax M2.5 / M2.5-Lightning

> An open-weight 229B MoE agentic model that runs frontier coding tasks at roughly one-twentieth the price of closed rivals.

- Maker: MiniMax
- Line: MiniMax M-Series
- Canonical page: https://ai-tldr.dev/models/minimax-m2-5/
- Updated: 2026-06-18

## In short

MiniMax M2.5 / M2.5-Lightning is an open-weight 229B MoE agentic model that runs frontier coding tasks at roughly one-twentieth the price of closed rivals. MiniMax released it on 12 February 2026. It has a 205K-token context window. You can use it through MiniMax, OpenRouter and Hugging Face.

## Facts

| Fact | Value | Source |
| --- | --- | --- |
| Context window | 205K (max output ~131K tokens) | [1] |
| Input price | $0.15 / 1M tokens (M2.5); $0.30 / 1M tokens (M2.5-Lightning) | [2] |
| Output price | $1.20 / 1M tokens (M2.5); $2.40 / 1M tokens (M2.5-Lightning) | [2] |
| Licence | Modified MIT (attribution required for commercial use) · open weights | [1] |
| Released | 12 February 2026 | [1] |
| Parameters | 229B total / 10B active (MoE) | [1] |
| Knowledge cutoff | Not disclosed | [1] |

## Sources

1. [Hugging Face model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)
2. [Official announcement — MiniMax M2.5](https://www.minimax.io/news/minimax-m25)

## Overview

MiniMax M2.5 is an open-weight large language model from Shanghai-based MiniMax, released on February 12, 2026. It uses a sparse Mixture-of-Experts design with roughly 229B total parameters but only about 10B active per token, which keeps inference cheap while preserving frontier-level quality on coding, tool use, web search, and office-document tasks. The model is text-only; MiniMax handles vision and audio through separate models.

The release shipped in two variants that share the same weights and capability but differ in speed: the standard MiniMax M2.5 and a faster M2.5-Lightning. MiniMax positioned both as built for real-world productivity, trained with reinforcement learning across many complex digital working environments rather than tuned purely for static benchmarks. It posts 80.2% on SWE-bench Verified and 76.3% on BrowseComp, and lands around 42 on the Artificial Analysis Intelligence Index — competitive with much larger closed models.

The headline pitch is price. At launch MiniMax priced M2.5 at roughly one-tenth to one-twentieth the cost of comparable closed frontier models such as Claude Opus 4.6, making sustained agentic use economically viable. The weights are published on Hugging Face under a modified MIT license that requires commercial users to display the 'MiniMax M2.5' name in their interface. It has since been succeeded in MiniMax's lineup by later M-series models (M2.7), and on the official API M2.5 is now listed among legacy models.

## Benchmarks

- SWE-bench Verified: 80.2% ([source](https://www.minimax.io/news/minimax-m25))
- Multi-SWE-Bench: 51.3% ([source](https://www.minimax.io/news/minimax-m25))
- BrowseComp: 76.3% ([source](https://www.minimax.io/news/minimax-m25))
- GPQA Diamond: 85.2% ([source](https://huggingface.co/MiniMaxAI/MiniMax-M2.5))
- AIME 2025: 86.3% ([source](https://huggingface.co/MiniMaxAI/MiniMax-M2.5))
- MMLU-Pro: 80.1% ([source](https://huggingface.co/MiniMaxAI/MiniMax-M2.5))
- Artificial Analysis Intelligence Index: 42% ([source](https://artificialanalysis.ai/articles/minimax-m2-5-everything-you-need-to-know))

## Pricing

- Input: $0.15 / 1M tokens (M2.5); $0.30 / 1M tokens (M2.5-Lightning) per 1M tokens
- Output: $1.20 / 1M tokens (M2.5); $2.40 / 1M tokens (M2.5-Lightning) per 1M tokens

Launch pricing from MiniMax. Both variants support prompt caching. M2.5 is now listed among legacy models on the MiniMax API; third-party providers (OpenRouter, SiliconFlow, Fireworks, Novita) may price it differently.

Source: https://www.minimax.io/news/minimax-m25

## Strengths

- Strong agentic coding: 80.2% on SWE-bench Verified, competitive with closed frontier models
- Very low cost per token relative to closed rivals, enabling long-running agent workloads
- Open weights under a permissive (modified MIT) license, downloadable and self-hostable
- Efficient MoE design — about 10B active parameters keeps serving cheap and fast
- Two serving tiers: standard M2.5 and a higher-throughput M2.5-Lightning variant
- Long 205K-token context window for large codebases and multi-step agent traces
- Built for real-world office and tool-use tasks, not just static benchmarks

## Best for

- Autonomous and semi-autonomous coding agents running across large codebases
- Cost-sensitive, long-horizon agent workflows where token volume is high
- Web-research and browsing agents (strong BrowseComp performance)
- Self-hosted or private deployment where open weights are a requirement
- Office and document-generation tasks (Word, Excel, PowerPoint workflows)
- Multilingual programming across Python, Go, C++, Rust, and TypeScript

## How to access

- [MiniMax](https://platform.minimax.io/docs/guides/pricing-paygo) — model id `MiniMax-M2.5`
- [OpenRouter](https://openrouter.ai/minimax/minimax-m2.5) — model id `minimax/minimax-m2.5`
- [Hugging Face](https://huggingface.co/MiniMaxAI/MiniMax-M2.5) — model id `MiniMaxAI/MiniMax-M2.5`

## MiniMax M-Series versions

- [MiniMax M3](https://ai-tldr.dev/models/minimax-m3/) (1 June 2026)
- [MiniMax M2.7 / M2.7-highspeed](https://ai-tldr.dev/models/minimax-m2-7/) (18 March 2026)
- MiniMax M2.5 / M2.5-Lightning (12 February 2026) — this page
- [MiniMax M2.1](https://ai-tldr.dev/models/minimax-m2-1/) (23 December 2025)
- [MiniMax M2](https://ai-tldr.dev/models/minimax-m2/) (27 October 2025)

## FAQ

### What is the difference between MiniMax M2.5 and M2.5-Lightning?

They share the same weights and capabilities and differ only in serving speed and price. The standard M2.5 runs around 50 tokens per second at $0.15/$1.20 per million input/output tokens, while M2.5-Lightning roughly doubles throughput to about 100 tokens per second at $0.30/$2.40 per million tokens.

### Is MiniMax M2.5 open source?

The weights are open and published on Hugging Face under a modified MIT license. It is open-weight: free to download and self-host, but the license requires that commercial users prominently display the 'MiniMax M2.5' name in their user interface.

### How many parameters does MiniMax M2.5 have?

It is a Mixture-of-Experts model with about 229B total parameters but only roughly 10B activated per token, which is what makes its inference unusually cheap for a model of its quality.

### Is MiniMax M2.5 still current?

No. M2.5 launched in February 2026 and has since been succeeded by later M-series releases (such as M2.7). On MiniMax's official API it is now listed among legacy models, though it remains downloadable from Hugging Face and available through third-party providers.

## Links

- [Official announcement — MiniMax M2.5](https://www.minimax.io/news/minimax-m25)
- [Hugging Face model card](https://huggingface.co/MiniMaxAI/MiniMax-M2.5)
- [MiniMax API pricing docs](https://platform.minimax.io/docs/guides/pricing-paygo)
- [Artificial Analysis — M2.5 analysis](https://artificialanalysis.ai/articles/minimax-m2-5-everything-you-need-to-know)
- [OpenRouter — MiniMax M2.5](https://openrouter.ai/minimax/minimax-m2.5)

Source: AI/TLDR LLM registry, https://ai-tldr.dev/models/minimax-m2-5/