MiniMax M2.5

SOTA in Coding and Agent, designed for Agent Universe

Designed for high-throughput, low-latency production environments. M2.5 delivers industry-leading coding and reasoning capabilities at a fraction of the cost.

Research

Performance Benchmark

Compared to its predecessor, M2.5 demonstrates greater decision-making maturity in handling Agentic tasks: it has learned to solve problems with more precise search iterations and better token efficiency.

M2.5 has achieved significant capability improvements in advanced workspace scenarios such as Word, PPT, Excel financial modeling, and more.

By combining reinforcement learning-optimized task decomposition with thinking token efficiency, M2.5 delivers significant advantages in both speed and cost when completing complex tasks.

M2.5 is available in 100 TPS and 50 TPS versions, with output pricing at just 1/10 to 1/20 of comparable models.

Coding Core Benchmark Scores Open-Source SOTA

M2.5 has reached the level of tier-one industry models. On the multilingual task benchmark Multi-SWE-Bench, M2.5 achieved the best performance in the industry.

M2.5 Showcases

See What M2.5 Can Do

All examples below were generated by MiniMax M2.5 in a single shot. Click cards to enlarge.

Cat Tunnel E-Commerce

E-commerce website for a premium modular cat tunnel system, featuring cinematic auto-play hero video, Scandinavian minimalist meets Wabi-Sabi aesthetics, with grid-based layout and architectural product photography

Strategy Consulting Presentation

Professional strategy consulting PPT for a sparkling beverage brand, covering non-cola carbonates market evolution, competitive landscape analysis, target consumer profiling, and distribution channel strategy

Interactive 3D Mountain Vista

Real-time 3D alpine landscape with photorealistic terrain rendering, drag-to-orbit camera controls, cinematic sunrise lighting, and dynamic cloud formations at adaptive resolution

Hogwarts Common Room Tour

Immersive 3D virtual tour of the Gryffindor Common Room with Gothic arched architecture, dynamic fireplace, floor-to-ceiling bookshelves, and multi-angle navigation including day/night cycle switching

Y2K Cyberpunk Music Visualizer

Retro-futuristic millennium-themed music player interface featuring neon cyan-magenta color palette, glitch grid aesthetics, real-time audio waveform visualization, and interactive playback controls

Rebate of 10% for the inviterDiscount of 10% for invitees

Invite friends, earn benefits

Subscribe to a Coding Plan got a 10% discount, while the inviter got a 10% rebate!

Rebate of 10% for the inviterDiscount of 10% for invitees

Invite friends, earn benefits

Subscribe to a Coding Plan got a 10% discount, while the inviter got a 10% rebate!

DEVELOPER TOOLS

Empowering Developer Choice

Outstanding Tool Scaffolding Generalization

01 / Access Method

Quick API Integration

Two API versions: M2.5 and M2.5-lightning with identical results but faster speed. Full automatic Cache support, no configuration needed.

PYTHONCURL

import requests

url = "https://api.minimax.io/v1/text/chatcompletion_v2"

payload = {

"model": "MiniMax-M2.5",

"messages": [

{"role": "user", "content": "Hello"}

]

}

headers = {"Authorization": "Bearer <token>"}

response = requests.post(url, json=payload, headers=headers)

print(response.text)

v2.1 API Connected

main.py

02 / Access Method

For AI Coding Tools

Model weights have been fully open-sourced on HuggingFace. It is recommended to use vLLM or SGLang for deployment to achieve optimal performance.

01 /

Subscribe to the Coding Plan

The price remains unchanged, while performance has significantly improved. Coding Plan users now automatically benefit from higher inference speeds.

02 /

Open Platform Integration

Supports standard M2.5 and high TPS version of M2.5-lightning. Coding Plan users will automatically receive a larger share of lightning resource allocation.

03 /

MiniMax Agent Integration

The general Agent platform based on M2.5 is now fully open. Experience the best programming assistance and logical reasoning capabilities without any development required.

03 /

Open Source and Local Deployment

We are committed to giving back to the community. M2.5 has been synchronously open-sourced on HuggingFace and GitHub, supporting private cluster deployment and fine-tuning.

Local Private Deployment

Model weights are fully open-sourced on HuggingFace. We recommend SGLang or vLLM for optimal performance, with Transformers and Ktransformers also supported.