MiniMax-M2.5

2.5

Open-weight MoE language model from MiniMax (229B params, 256 experts) for agentic coding, reasoning and tool use; 192K context window.

✓ Active✓ Public access⚖ Open weightsLLMReasoning modelTool-using model

Context window

192K

tokens

Parameters

229B

parameters

Release date

12 February 2026

🏢MiniMaxProducer

Access:APIDownloadDeployment:☁ Cloud💻 Local

Overview

MiniMax-M2.5 is an open-weight large language model (LLM) developed by the Chinese company MiniMax and released in February 2026 on Hugging Face. It uses a Mixture-of-Experts (MoE) architecture with 256 experts, of which 8 are activated per token, together with a multi-token prediction (MTP) mechanism.

The model targets agentic workloads: coding, multi-step reasoning and tool use / function calling. It supports a context window of 196,608 tokens (about 192K) and operates on text only. The weights are available for download under a Modified MIT license.

According to the model card, MiniMax-M2.5 reaches 80.2% on SWE-Bench Verified, 85.2 on GPQA-Diamond and 86.3 on AIME25. Besides downloadable weights, the model is also offered through the MiniMax API.

Classification

LLMReasoning modelTool-using model

Access & deployment

APIDownload

CloudLocal

Weights: Open weights

Key parameters

📏 Context: 192K

🧩 Parameters: 229B

✓ Tools

📥 Input: text

Technical specification

Context window

192K

tokens

Parameters

229B

parameters

License

Modified MIT

Features:✓ Tool use

Modalities

⬇ Input

text

⬆ Output

textcode

Capabilities and applications

Native model capabilities

Reasoning

The model's ability to reason logically and solve complex problems.

Category: reasoning

Multi-step reasoning

Carrying out multi-step chains of reasoning across long, complex tasks.

Category: reasoning

Long context

Support for large context windows — tens to hundreds of thousands (or millions) of input tokens. Enables analysis of entire codebases, long documents, and many parallel conversations without losing earlier information. GPT-5.1 supports 400,000 tokens.

Category: language

Coding

Generating, analysing and modifying code in many programming languages. Covers writing functions, debugging, refactoring, code review, and creating tests. Measured by benchmarks such as HumanEval and SWE-bench.

Category: coding

Function Calling

Category: planning

Structured output

Producing data in structured formats such as JSON.

Category: structured_generation

Multilingual

Competence in many natural languages (from a few to over a hundred): understanding, generation, translation, and code-switching within a single conversation. Frontier models support a wide range of languages with comparable quality.

Category: language

Planning

Forming and executing action plans for complex tasks.

Category: planning

Streaming output

Category: reasoning

Benchmark results

5 benchmarks

SWE-Bench Verified

accuracy

80.2%