I Let a Chinese AI Control My Computer: Minimax M3 Tested

I handed control of my computer to a Chinese AI model. Not a chatbot. Not a coding assistant. A system that moves my mouse, clicks my buttons, and completes tasks while I watch. The model is Minimax M3, and after running it through real e-commerce research and profit-margin calculations, I can tell you this: the economics of AI automation just changed dramatically.

Key Takeaways

• Minimax M3 outperformed GPT 5.5 on real software task benchmarks (59% vs. 58.6%) and browser research tasks (83.5% vs. Claude Opus)
• Costs roughly 20x less than GPT 5.5 — approximately $0.60 per million tokens
• First open-source model with true “computer use” — it controls your desktop, mouse, and keyboard autonomously
• 1 million+ token context window handles entire projects and documents without forgetting
• Multimodal understanding processes text, images, and video in a single model
• Not perfect on every task — browser automation especially burns through tokens and time

What Makes Minimax M3 Different From Other AI Models

Minimax is already one of China’s most powerful AI companies, known primarily for their video generation system. But the M3 release shifts the conversation entirely. This is the first open-source model to combine three capabilities that were previously scattered across expensive, closed systems.

First, the coding ability is genuinely top-tier. Second, that 1 million+ token memory means I can feed it entire project folders, long documentation, or massive spreadsheets without the model losing track of earlier context. Third — and this is the critical part — it understands text, images, and video together in one unified system.

But the feature that made me download it immediately is computer use. M3 doesn’t just suggest code or answer questions. It can literally operate a desktop computer: watching the screen, moving the mouse, clicking, typing, and completing multi-step workflows. I’ve tested AI coding assistants before. I’ve tested browser agents. This is the first time I’ve used an open-source model that does both while sitting inside my actual operating system.

The Benchmark Evidence: Where It Actually Beats GPT 5.5

I know benchmark claims are cheap. So I went to the actual numbers. On SWE-bench Verified — the benchmark for real software engineering tasks — Minimax M3 scored 59%. GPT 5.5 scored 58.6%. Google’s Gemini 2.5 Pro managed 54.2%. These are narrow margins, but they’re margins in favor of a model that costs a fraction of the price.

More impressive to me was BrowserComp, the test for independent web research and navigation. Here M3 hit 83.5%, surpassing even Claude Opus. This matters because browser automation is where I burn the most money in my business — and where most AI agents fail. The model also ranked at the top of OSWorld tests for general computer use and terminal operation.

Let me put this in perspective: an open-source Chinese model is challenging the world’s most expensive closed systems on their own benchmarks, while costing approximately $0.60 per million tokens. GPT 5.5-level performance at grocery-store prices.

My Real Test: E-Commerce Research and Profit Calculations

Benchmarks are useful. Real tasks reveal the truth. I gave M3 two assignments directly relevant to my e-commerce work.

Test 1: eBay UK Top-Seller Research

I opened Minimax’s Agent mode and typed: “Research the top 5 best-selling products on eBay UK, find average prices, and give me a one-page summary report.”

The system immediately created a plan and started executing. It opened the browser, navigated to eBay, searched categories, and compiled data. The entire process took roughly a minute. What struck me wasn’t just the speed — it was the stability. I didn’t wait an hour. I didn’t get a timeout error. The agent completed the research loop and presented formatted results.

Here’s what most people miss: browser automation is token-expensive. When M3 scrolls, clicks, encounters an error, retries with a different approach — every action consumes tokens. I’ve seen tasks that would cost 1 million tokens in a clean coding session burn far more when the model is navigating a live website. The low per-token cost becomes essential here.

Test 2: Automated Profit-Margin Calculator

For my second test, I asked M3 to build a working tool: “Create a single-page HDMI cable product calculator. Calculate cost, sale price, eBay fees, PayPal conversion fees, and net profit. Turkish interface.”

Within seconds, it generated a functional calculator. I tested it immediately: product cost 100 TRY, sale price 50 TRY. The system automatically applied 12% eBay commission (correct for that category) and PayPal’s 2.9% + $0.30 fee structure. Final calculation: 57.75 TRY net. Then it flagged the obvious: “You’re losing money at this price.”

This is functionality I’ve paid subscription fees for — tools costing $100+ monthly for eBay and Amazon sellers. M3 built a working alternative in seconds, customized to my specifications, running locally.

The Honest Downsides Nobody’s Talking About

I need to be direct here. M3 is new. Very new. And not every task runs perfectly.

During my YouTube channel search demo, the model hit a page resize error, had to backtrack, and try alternative navigation paths. The task completed, but the retry loops add time and token consumption. For complex browser tasks, you might wait several minutes. For simple coding tasks, it’s nearly instant.

I also noticed that Thinking Mode (the deeper reasoning setting) produces noticeably better results for research tasks. Running M3 without this optimization gives faster but less reliable output. There’s a trade-off between speed and accuracy that you’ll need to calibrate for your own workflows.

Most importantly: this is not a magic money machine. I don’t make income guarantees. The model makes mistakes. It misreads interfaces. It can get stuck in loops. What it offers is capability at a price point that changes the math for small operators like me.

How to Actually Use Minimax M3 Today

I’ve tested three access methods, and all are functional:

Desktop Application: Download Minimax directly. Select your model, type your task, and choose whether the agent gets full computer control or asks permission at each step. I recommend starting with permission-required mode until you understand its behavior.

OpenRouter Integration: For developers already using unified API interfaces, M3 is available in the chat section. Select the model and use it identically to other providers.

Minimax Agent Platform: This is where it gets interesting. The platform includes pre-built agents like Hermes and MaxClaw, plus the ability to create agent teams — one codes, one reviews, one oversees quality. You can add skills from the marketplace, connect your phone for mobile tasks, and upload documents or images as project context.

I run M3 in Thinking Mode for research tasks, use agent teams for multi-step quality control, and keep a close eye on token consumption during browser automation.

What This Actually Means for Online Business Owners

I’ve built four companies. I’ve spent tens of thousands of dollars and years of development time creating automation systems that M3 now approximates with a single sentence prompt. The direction is unmistakable.

But I want to correct a common misconception. AI doesn’t take your job. People using AI take jobs from people not using AI. The competitive advantage isn’t the technology itself — it’s the operator who learns to deploy it effectively.

For my own service business, we now offer clients 24/7 AI agent systems that optimize workflows continuously. The businesses adopting these structures will outpace those relying on manual processes. Not because AI is perfect, but because good enough automation at near-zero marginal cost reshapes entire market segments.

A year ago, building these capabilities required engineering teams, substantial budgets, and months of development. Today, a Chinese open-source model costing pennies per million tokens sits on my desktop and executes tasks while I review the output.

Frequently Asked Questions

Is Minimax M3 really free to use?

The model weights are open-source, but API and hosted usage have costs. Through OpenRouter or Minimax’s platform, pricing runs approximately $0.60 per million tokens — roughly 20 times cheaper than GPT 5.5 equivalent usage. Desktop app installation is free.

How does computer use AI differ from regular chatbots?

Standard chatbots process text and return text. Computer use AI perceives your screen visually, controls mouse and keyboard inputs, navigates applications, and completes multi-step workflows across different software. It operates software rather than just describing how to use it.

Can Minimax M3 replace my existing e-commerce tools?

For specific tasks like research, calculation, and content generation, yes — with caveats. It won’t match specialized tools for inventory management or accounting integrations. I use it to supplement rather than fully replace established workflows, especially where those tools charge substantial monthly fees.

Is it safe to let an AI control my computer?

Minimax offers permission modes where the agent pauses for approval before each action. I strongly recommend starting with this setting. Full autonomous mode carries risks of unintended clicks, data exposure, or task execution errors. Treat it like giving remote desktop access — useful but requiring oversight.

Conclusion

Minimax M3 isn’t flawless. It’s not a guaranteed income system. What it represents is more important than any single task it completes: capable AI computer use at prices that make experimentation virtually free. The benchmark victories over GPT 5.5, the 20x cost advantage, the open-source accessibility — these combine to accelerate how quickly solo operators and small teams can automate complex workflows.

I handed my computer to a Chinese AI. It researched eBay markets, built profit calculators, and warned me when my pricing would lose money. The task took minutes. The cost was negligible. Whether this excites or concerns you probably predicts how you’ll fare in the next phase of online business automation.

Watch the full video (in Turkish — English subtitles available):

Tools & Community

• TurkoLister — the AI listing tool I use to turn Amazon products into optimized eBay UK listings in about 60 seconds (from £4.99/month, £1 one-week trial).
• AI & E-commerce Community — my Turkish-speaking community ($19/month) with weekly live sessions.
• Subscribe on YouTube — new experiments every week.