Minimax
MiniMax M2.5
MiniMax-M2.5
Total Context
204.8K
Max Output
131.1K
Released
N/A
Input$0.3 / M tokens
Output$1.2 / M tokens
MiniMax M2.7
MiniMax-M2.7MiniMax M2.7 is presented as a productivity and engineering model for autonomous workflows, multi-agent collaboration, live debugging, and document-heavy work. Public descriptions mention root-cause analysis, financial modeling, and full Word/Excel/PowerPoint-style document generation. It should be described as an applied work model, not just a chat or writing model.
Context Window
204.8K tokens
Maximum Output
131.1K tokens
Release Date
Mar 18, 2026
Modalities
MiniMax M2.7 Pricing
| Input Price | Output Price | Cache Read |
|---|---|---|
| $0.3/M | $1.2/M | $0.06/M |
MiniMax M2.7 API Capabilities
Reasoning
Supported
Tool calling
Supported
Temperature parameter
Supported
Attachments
Not supported
Knowledge Base
—
Endpoint Protocols
Completions APIMessages APIgemini
MiniMax M2.7 Model Highlights
MiniMax M2.7 is a text reasoning model focused on real-world software engineering, coordinated agent work and editable professional deliverables.
System-Level Engineering
Correlates logs, monitoring metrics, traces and database evidence to diagnose defects and make system-level engineering decisions.
Coordinated Agent Teams
Supports agent teams with stable roles and autonomous decisions for dividing and coordinating complex multi-part work.
Professional Document Editing
Performs high-fidelity, multi-round work on Word, Excel and presentation materials while preserving editable output.
MiniMax M2.7 Use Cases
MiniMax M2.7 fits production incident investigation, collaborative engineering projects and multi-round preparation of professional office materials.
Production Incident Response
Review logs, traces, metrics and database state to identify a likely root cause and prepare a scoped recovery or remediation plan.
Multi-Agent Engineering
Assign analysis, implementation and verification to coordinated agent roles, then combine their outputs into one engineering result.
Editable Office Deliverables
Draft and repeatedly refine reports, spreadsheets or presentations while preserving structure and producing material suitable for further editing.
How to Use MiniMax M2.7 via the TokenHub API
import OpenAI from "openai"
const client = new OpenAI({
apiKey: process.env.TOKENHUB_API_KEY,
baseURL: "https://us-api.tokenhub.com/v1",
})
const result = await client.chat.completions.create({})
console.log(result.choices[0]?.message?.content)MiniMax M2.7 Benchmarks
MiniMax-M2.7
| Index score | ||
|---|---|---|
| Artificial Analysis Intelligence Index | Artificial Analysis broad capability aggregate | 38.1 |
| Artificial Analysis Coding Index | Artificial Analysis software task aggregate | 41.9 |
| Knowledge & Reasoning | ||
| GPQA | Advanced science problem solving | 87.4% |
| HLE | Broad expert-level exam set | 28.1% |
| Coding & Engineering | ||
| SciCode | Scientific coding challenges | 47% |
| Terminal-Bench Hard | Hard terminal task execution | 39.4% |
| Instruction Following & Agent Tasks | ||
| IFBench | Prompt constraint adherence | 75.7% |
| AA-LCR | Long-context reasoning | 68.7% |
| τ²-Bench | Agent workflow tasks | 84.8% |
Metrics sourced from Artificial Analysis
MiniMax M2.7 Guides & Articles
MiniMax M2.7 FAQ
MiniMax M2.7: capabilities, use cases, limits, and TokenHub guidance.
What does MiniMax M2.7 focus on?+
MiniMax M2.7 is a MiniMax model for real-world software engineering and agent delivery.
Which projects fit MiniMax M2.7?+
Best for iterative engineering delivery, debugging and refactoring and repository-scale development, especially when long-horizon task completion is the priority.
What is special about MiniMax M2.7?+
Key strength: strong real-world engineering, debugging, and end-to-end delivery.
When is another model better?+
Long autonomous runs can consume substantial time and tokens. For multimodal input, consider MiniMax M3.
How do I avoid ID mistakes?+
Use the exact ID shown by TokenHub; follow your account docs and verify current features.
Ready to use MiniMax M2.7?
Use one API key to access MiniMax M2.7 and more AI models through TokenHub.
Media and Discussions
Selected public videos and posts related to this model.
X (Twitter)
MiniMax (official) ·
https://t.co/YkxS1M2CdV
View originalOpenRouter ·
MiniMax M2.7 from @MiniMax_AI is live on OpenRouter! M2.7 sees a large jump in agentic and tool calling capabilities. https://t.co/pOXGg0VnR1
https://pbs.twimg.com/media/HDtsxTra4AAZ8an.jpgView originalArtificial Analysis ·
MiniMax-M2.7 achieves one of the highest scores in GDPval-AA (agentic real-world work tasks) with an Elo of 1495. This places it ahead of GPT-5.3 Codex (xhigh, 1462) and alongside GPT-5.2 (xhigh, 1462) https://t.co/pxTjMlgNVn
https://pbs.twimg.com/media/HDtUQXlaUAQpmLe.jpgView originalReddit
MiniMax M2.7 on OpenRouter
iamn0 ·
View originalMy first impressions of using MiniMax M2.7 as my new main agent
markusla ·
Read full post
There was a announcement yesterday about the new MiniMax M2.7 Model. Introducing MiniMax-M2.7, our first model which deeply participated in its own evolution, with an 88% win-rate vs M2.5 - Production-Ready SWE: With SOTA performance in SWE-Pro (56.22%) and Terminal Bench 2 (57.0%), M2.7 reduced intervention-to-recovery time for online incidents to 3-min on certain occasions. - Advanced Agentic Abilities: Trained for Agent Teams and tool search tool, with 97% skill adherence across 40+ complex skills. M2.7 is on par with Sonnet 4.6 in OpenClaw. - Professional Workspace: SOTA in professional knowledge, supports multi-turn, high-fidelity Office file editing. I've been a huge user of Opus and Sonnet as my main agent for all orchestration and every time I've switched to GPT 5.4, GPT 5.4-Pro, Gemini or any other model, I've cursed and pulled my hair. I've each time thought that if this is what people experience when trying openclaw with any of these models, i understand they think its useless and sucks. The difference in level using Opus and Sonnet has been astronomical, but it's equally expensive... So I've been searching for an alternative. And when I saw that announcement I of course gave it a try, with not huge hopes as I was really dissapointed with MiniMax 2.5 and felt it was quite useless for anything advanced or tooling etc. With the pricing of MiniMax being ~10x the cost of Anthropic it does feel very interesting to switch, if its good enough. Anyway, I setup M2.7 and instructed my Openclaw to setup some Benchmarks using both Coding and Tooling, comparing MiniMax M2.7 with GPT 5.4, Gemini 3.1 Pro and some others. Actually, MiniMax came out on top. It was the fastest to deliver a working result. I then deviced some of my own Tooling challenges that I know the models often struggle with. In the lines of; Connect to system X (find IP, credentials etc), for the process Y grab the config file (which it had to do sudo to access) and compare it with another similar file on my local system and report the diff. So a multi step tool chain which some of the models totally failed while the most advanced succeeded with. MiniMax M2.7 succeeded there too, and was the fastest. That gave me the confidence to try it as my main model, and now ~5 hours of active usage, lots of tooling use and troubleshooting some systems (but no coding done, yet!). I've not missed Sonnet or Opus once. The ones who's tried, what's your impression so far?
MiniMax-M2.7 Announced!
Mysterious_Finish543 ·
View originalYouTube
Let's Run MiniMax M2.7 - #1 Coding Local AI for Agents & OpenClaw? 🧐
xCreate
View originalMinimax M2.7 Is INSANE - Fully Testing A SELF Evolving Model!
Bijan Bowen
View originalMinimax M2.7: Self-Evolving Agent Model
Developers Digest
View original