"The Impossible Triangle" was shredded.
Just this past weekend, the AI community blew up - MiniMax quietly went live with a new generation of open source big models MiniMax M2Not only did it break into the world's top five in terms of performance, but it also knocked the price right down to Claude's 4.5 8%furthermoreFully open source and free for a limited time!
What's even more dramatic is that it's fully blossomed in multiple dimensions such as programming, Agent workflows, deep search, multimodal tasks, etc., and the real-world performance evenMilling quite a few closed-source top-flow modelsToday, we're going to take apart the MiniMax M2, the "light of China". Today, we will deeply disassemble this "light of China" - MiniMax M2, in the end, what is strong? How should ordinary people use it?
What is 🔥 M2? Parameters, Architecture, Positioning
MiniMax M2 is the next generation of MiniMax.hybrid thinking model(Mixture-of-Thoughts), a staple of Coding + Agentic Workflows(Intelligent Body Workflow) Scenario.
- Total number of participants:: 230B
- Number of activation parameters:: 10B only (efficient reasoning, cost-effective pulling)
- open source status: Fully open source at Hugging Face
- free strategy: API and Agent platformsFree until November 7The CEO himself promised, "Until the servers can't hold."
- inference speed: TPS over 100 tokens/secondIt is far superior to similar domestic models such as DeepSeek.
- prices: $2.1/million tokens input, $8.4/million tokens output -- less than 1/12th of Claude's!
The bottom line is that M2 Not "toy models."Instead, they are "all-rounders" who can run full-stack development, in-depth financial research, 3D rendering, and multimodal tours.
🧠 M2's core competency: not just writing code, but the "brain of an intelligent body"
1️⃣ Programming ability: solidly in the top tier globally
In the authoritative list SWE-bench Verified On, M2 score 69.4which is second only to GPT-5 and Claude 4.5.number one in China!
Real-world case:
Knowledge card generationClaude 4.5 just "won't open": given a link to Zhihu, M2 not only summarizes the content, but also designs a beautiful, responsive front-end page.
3D Particle Galaxy:: One sentence. Create a 3D particle galaxy with swirling nebulas...While M2 successfully renders dynamic nebulae, Claude only outputs a snowflake.
Full Stack Merit Box Website: Users only need to say "click +1 merit, save to database, with administrator backend", M2 automatically outputs the complete front-end and back-end + database + admin interface, and then M2 will automatically output the complete front-end and back-end + database + admin interface.Not a single line of code to writeThe
In a color memory game live test, the author couldn't stop playing: "I went straight to level 7!"
2️⃣ Agentic capabilities: true "mission enforcers"
M2 doesn't just answer questions, it canProactive tool mobilization, planning steps, closed loop implementationThe
Renovation Budget Program: Upload a 90 sq ft house plan of Beijing, M2 auto-generated Simple modern style budget table within 150,000(with areas, materials, unit prices, remarks), also with a sketch of the layout (e.g. of a building)!
Financial Literature Review: Fuzzy command "I want to write a review of gold price fluctuations", M2 auto-retrieve 100+ Authoritative Literature(with Chicago Fed, JEB, JPMorgan Chase reports) with authentic citation source -Farewell to the "illusion paper."The
Anonymous Wishing Wall: Generate complete web applications in a single sentence, with automatic connection to Supabase, deployment of front-ends, and support for real-time interactions.out-of-the-boxThe
🛠️ How to use M2? Zero Threshold Hands-on Guide
✅ Free experience portal (domestic/overseas support):
- MiniMax Agent (no code required)
Domestic version:https://matrix-pre.xaminim.com/
Overseas Edition:https://agent.minimax.io/ - API calls (developer friendly)
Open Platform:https://platform.minimaxi.com
Model Download (Hugging Face):https://huggingface.co/MiniMaxAI/MiniMax-M2
✅ Access to mainstream IDEs such as Claude Code:
Simply add the ~/.claude/settings.json Configure the MiniMax API Key and model name in the MiniMax-M2The program can be called directly from Cursor, Claude Code and other tools.Experience silky smoothness like a native modelThe
✅ Self-deployment (suitable for businesses/researchers):
- Recommended vLLM or SGLang
- Memory Requirements: >220GB (supports 400,000~3,000,000 token contexts)
- courtesyChinese Deployment Documentation, hand in hand, honestly pulling it all together.
💡 Why does M2 break the "impossible triangle"?
In the field of large models, there has always been an "impossible triangle":
- high performance(INTELLIGENCE)
- low prices
- high speed
In the past, you could only choose two out of three. But M2 took all three on its own:
| dimension (math.) | M2 Performance |
|---|---|
| intelligence level | 5th globally, 1st in open source (Artificial Analysis) |
| prices | Claude's 8%, one of the lowest in the country. |
| tempo | 100+ TPS, far more than DeepSeek, GLM |
One sentence summary M2::
"A compact, fast and superbly cost-effective model built for Coding and Agent."
🌟 Write in the end
The MiniMax M2 is not just a technological breakthrough, it is aThe Revolution of Inclusive AIThe
Previously, top AI was a "supercar" - a powerful car that ordinary people couldn't afford or wouldn't drive.
Now, the M2 is a "fully autonomous smart car" - you name the destination, it'll take you there.It's also fuel efficient, free, and traffic freeThe
The free window is limited, so hurry up before the servers "can't hold out"!
🔗 Experience Address:
👉 https://agent.minimaxi.com
👉 https://platform.minimaxi.com