MiniMax 2.7 Goes Open Source: Can Chinese Models Finally Code?

Chinese large models have long struggled with coding capabilities. While they perform well in Chinese comprehension and dialogue generation, writing code often exposes their weaknesses—syntax errors, logic bugs, context misunderstanding, you name it.

But the release of MiniMax 2.7 shows signs of change.

According to data released by Xiyu Technology, MiniMax 2.7 scored 56.22% on the SWE-Pro coding benchmark. To put this in perspective: Claude Opus’s latest version is around 57%, GPT-4 Turbo sits at roughly 54%. In other words, MiniMax 2.7’s coding ability is approaching the international first tier.

I tested it myself. I asked it to write a simple Python web scraper to fetch news headlines. The results surprised me—clean code structure, proper exception handling, and decent comments. It missed one detail (no headers, leading to anti-scraping blocks), but overall quality exceeds many junior programmers.

More notably, MiniMax 2.7 also shows significant improvement over its predecessor in OpenClaw and MMClaw evaluations. This suggests the improvement isn’t just “score chasing” but genuine capability enhancement.

Of course, I found some issues too. When handling complex algorithms, it still makes basic mistakes like incomplete boundary condition handling. And for multi-language support, its understanding of niche languages like Go and Rust clearly lags behind Python and JavaScript.

Overall, MiniMax 2.7 going open source is good news for domestic developers. At least we have another reliable domestic option when choosing coding assistants. Plus, open source means local deployment— a major plus for enterprises concerned about data security.

It’s good to see Chinese large models catching up on coding capabilities. This is a promising start.