01What Is Minimax M2.5 and How It Works

Minimax M2.5 is an AI model from the Shanghai startup Minimax released in early 2026. It scores 80.2% on the SWE Bench Verified test, which measures how well AI models solve real software engineering tasks. That is only 0.6% below Claude Opus 4.6 from Anthropic, which scored 80.8%.

The big difference is cost. M2.5 costs 95% less to run than Claude Opus. Julian Goldie explains that this model makes frontier intelligence "too cheap to meter." Dollar per hour, it runs continuously at 100 tokens per second for a fraction of the price.

The key to its performance is the Forge Framework. Instead of learning from human feedback after answering questions, M2.5 is dropped into 200,000 real-world coding environments. It learns by doing, like fixing actual bugs in real GitHub repositories across Python, Rust, Go, TypeScript, C++, and seven other programming languages.

02Main Features of Minimax M2.5

Minimax M2.5 has several standout features. First, it uses only 230 billion total parameters, with just 10 billion active at any time. This makes it extremely efficient compared to models that use hundreds of billions of parameters all at once.

Second, it acts like a software architect. Before writing code, M2.5 plans the structure, maps out features, and designs the interface. This behavior emerged during training, not from programming.

Third, it performs well on multiple benchmarks:

  • Multi-SWE bench: 51.3% vs. Opus 4.6 at 50.3%
  • Function calling: 76.8 on the Berkeley leaderboard, higher than Claude, GPT 5.2, and Gemini 3 Pro
  • Average task time: 22.8 minutes per SWE bench task, nearly identical to Claude's 22.9 minutes

Finally, M2.5 is open source with full model weights on Hugging Face under a modified MIT license. You can download and deploy it on your own servers.

03How to Set Up Minimax M2.5

  1. Download the model weights from Hugging Face. Minimax released the full model under a modified MIT license, so you can get it for free.
  2. Choose your deployment tool. Julian says you can use VLM or SG Lang to run the model on your own servers.
  3. Integrate with coding tools. M2.5 already works with Cursor, Client, Kilo Code, Claude Code, and OpenHands.
  4. Set up an API key if you prefer to use Minimax's API. The cost per task is about 15 cents.
  5. Test it for free on several platforms that offer limited-time free access. Julian recommends trying it this week.
  6. Run continuous instances. Four M2.5 instances running 24/7 for a year costs under $10,000. That is less than a junior developer's monthly salary.

04How to Use Minimax M2.5 for Coding Tasks

Once you have Minimax M2.5 set up, using it is straightforward. Julian explains that the model plans before it codes. When you give it a problem, it first maps out the structure, designs the interface, and then writes code.

For real-world bugs, M2.5 works in actual GitHub repositories. It can handle multi-file code bases across Python, Rust, Go, TypeScript, C++, and seven other languages. Sixteenth examples show it fixing bugs in minutes.

For function calling, M2.5 chains together multiple tools in sequence. This is important for autonomous agents that need to search, analyze, and take actions. On the Berkeley leaderboard, it scores 76.8, beating Claude and GPT 5.2.

For continuous operation, you can run the model 24/7. Minimax uses M2.5 internally for 30% of all tasks and 80% of newly committed code. Julian suggests using it for coding, searching, document analysis, PowerPoint creation, and financial modeling in Excel.

05Results and Comparisons Between Minimax M2.5 and Claude Opus

Julian shares several side-by-side comparisons between Minimax M2.5 and Claude Opus 4.6. On SWE Bench Verified, M2.5 scores 80.2% versus Opus at 80.8%. That is a gap of only 0.6%.

On the multi-SWE bench, which tests performance across multiple code bases at once, M2.5 actually beats Opus. It scores 51.3% compared to Opus's 50.3%. On function calling, M2.5 scores 76.8 on the Berkeley leaderboard, higher than Claude, GPT 5.2, and Gemini 3 Pro.

Cost is where M2.5 dominates. The cost per SWE bench task is roughly 15 cents for M2.5 versus about $3 for Claude Opus. That is a 20X difference.

Task completion speed is nearly identical: M2.5 averages 22.8 minutes, while Claude Opus averages 22.9 minutes.

The Open Hands team independently tested M2.5 and called it the first open-weight model to match Claude Sonnet's quality. Kilo Code found it achieving 80.2% out of the box, while Opus needs specific prompt modifications to push past 80%.

06Tips, Limits, and Mistakes to Avoid With Minimax M2.5

Julian notes several important points about Minimax M2.5. First, benchmarks aren't everything. Self-reported numbers always need scrutiny.

M2.5 is impressive on tests, but real-world use may vary.

Second, Minimax is still burning cash. In the first nine months of 2025 alone, they lost $512 million. They are in what CEO Yan Yunji calls a "nascent stage of monetization." There is a gap between benchmark performance and profitability.

Third, don't ignore the speed of change. Minimax released M2.1 in January 2026 at 74% on SWE Bench. Five weeks later, M2.5 hit 80.2%.

That is a six-point jump in a month. American labs took a year for similar gains.

Finally, Julian warns that the bottleneck is no longer technology—it is imagination. The real question is how to manage hybrid teams of humans and AI agents. Start piloting now, not next quarter, because competitors in Asia are already deploying this.

07Who Minimax M2.5 Is For and Why It Matters

Minimax M2.5 matters for several groups. For startup founders, it means running frontier-level AI agents continuously for the cost of a Netflix subscription. Four instances running 24/7 for a year costs under $10,000.

That replaces a team of AI workers for less than one junior developer's salary.

For managers at mid-sized companies, this is the moment to pilot AI agents. Julian says competitors in Asia are already deploying M2.5. It works in Cursor, Claude Code, and other tools today.

For developers, M2.5 is open source and free to try. You can swap it in today and see immediate results. Minimax itself uses it for 30% of all tasks and 80% of newly committed code.

For policy leaders, M2.5 shows that export controls did not stop Chinese AI. Instead, they forced Chinese labs to become more efficient. The model is smaller, cheaper, and faster, and within a rounding error of the best models on Earth.

08The Geopolitical Impact of Minimax M2.5

Julian highlights a deeper issue with Minimax M2.5. The United States spent three years trying to block Chinese AI through chip export controls. But Minimax built a frontier model with only 10 billion active parameters, matching models trained on clusters costing hundreds of millions of dollars.

The company was founded in December 2021 by Yan Yunji, a former vice president at SenseTime. They raised $600 million from Alibaba at a $2.5 billion valuation. In January 2026, they went public on the Hong Kong Stock Exchange, with shares doubling on the first day of trading.

Minimax has over 212 million users across 200 countries. Revenue grew from $3.5 million in 2023 to $30 million in 2024, and $53 million in the first nine months of 2025. 70% of that revenue comes from outside China.

The real takeaway: intelligence is becoming a commodity. When the best AI costs a dollar an hour, the bottleneck is imagination. Julian says the question is not whether to engage, but whether you do it on your terms or someone else's.