I Tried Every Chinese AI Model. Here's My Honest Take as a Non-Engineer.

I've spent the last couple months testing basically every major Chinese open-source AI model: DeepSeek, Kimi, and now MiniMax. The only one I haven't really gotten my hands on yet is the latest Qwen. I want to share what I found, because I think my perspective is different from most of the reviews you'll read online — I'm not an engineer. I'm just a guy who's been trying to use these models to actually get stuff done.

I'll be blunt: DeepSeek and Kimi were really disappointing for me.

I was using both of them as the reasoning engine inside OpenClaw, basically the brains of an AI agent that's supposed to figure out tasks and execute them. And they just couldn't handle it. They don't deal with ambiguity well at all. If you give them a clear, specific instruction, they're fine. But the moment something is even slightly vague or requires the model to figure out what you probably meant, they fall apart. They'd loop on the same task, make weird decisions, or just give up entirely.

Now, I know there are actual software engineers out there who swear by DeepSeek. And I believe them. If you're really good at writing precise, specific prompts and you know exactly what you want the model to do at every step, you can probably make it work. But that's the thing — as a non-technical person, I shouldn't have to be an expert prompt engineer to get basic results. That's the whole point of these models being "smart."

Then I tried MiniMax M2.7. And honestly? It's been solid.

I've been using it for about four days now on the $10/month token plan, running it as the main model in OpenClaw. It handles ambiguity way better than DeepSeek or Kimi did. When I give it something that's not perfectly spelled out, it actually reasons through what I probably want instead of just freezing or looping. It's not perfect — the grammar gets a little weird sometimes, and it has its dumb moments. But so does Opus 4.6, and that costs way more.

For $10/month, the value is kind of insane. I've been pushing it pretty hard and I haven't run out of tokens yet. If you're running an AI agent on a budget and you're not a hardcore developer who can squeeze perfect behavior out of DeepSeek with surgical prompting, MiniMax is the move right now.

Maybe the engineers who read this will disagree with me, and that's fine. I'm just sharing what worked and what didn't for someone who's learning as they go. If you've had a different experience with any of these models, I'd be interested to hear about it.