After coding with MiMo vibe over the weekend...

After MiMo was released, I tried it out using the free credits on OpenCode to do a small project. The project itself wasn’t very difficult, but MiMo froze for a long time and didn’t give any clear error feedback. At that time, I had a very bad impression of MiMo.

However, on Friday night, I saw someone say that the token billing method on MiMo’s official website is much more effective than the free credits on OpenCode. So I recharged 200 yuan and continued working on my original project, as well as two slightly more complex data cleaning tasks.

My feeling is that it is indeed much stronger than all the domestic models I used with the Coding Plan before.

This may not be reflected in its success rate per single execution, but I previously sensed that domestic large model vendors tend to impose some restrictions on agent calls: perhaps limiting the length of the thought chain or the number of agent rounds, causing those agents to tend to stop at around 50% to 60% of the task completion. For truly difficult corner cases or complex bugs, they tend to ignore them.

With MiMo, I think this problem is much better. It can indeed run solidly for several hours just to solve some tricky bugs. The logic is quite easy to understand: with the Coding Plan, the more calls you make, the higher the cost; but with token billing, the more calls, the more revenue.

In my actual usage, although I spent 200 yuan, I solved three tasks that had been bothering me for quite a while. I personally think it’s quite worth it. Even if I bought ready-made data online, it would definitely cost more than 200 yuan.

But this experience made me think of a contradiction with domestic models:

For example, Claude’s top-tier models can now replace a large part of the work, and domestic model vendors are actually capable of reaching 80% to 90% of Claude’s level. However, if the promotion continues mainly through the Coding Plan approach, the actual user experience will still be quite poor. Slightly more complex or tricky tasks cannot be handled, not because the models lack capability, but because of restrictions imposed by the vendors on calls. These cost-related restrictions, in turn, affect the adoption and promotion of agents in practical work.

I think this problem mainly stems from computational power limitations and the pricing habits of domestic service-based models. I wonder what everyone thinks about this?

My personal view is that in the next 3 to 5 years, there will still be a huge demand for computing power, but the real question is whether it will benefit companies like NVIDIA or truly promote the upgrade of domestic chips.

(Recently, some people also said that DeepSeek V4 has been delayed because it needs to adapt to domestic chips, which has caused training to not converge well…)