How to Run OpenClaw with Open-Source Models

I started off testing out OpenAI’s GPT-5.4, but encountered challenges with the model being lazy. I would, for example, ask it to perform a task I know it’s able to do, and experienced the model simply giving up after a few attempts.

This is, of course, unacceptable when it comes to a helpful assistant, so I decided to start trying out other alternatives, and came across a suite of Chinese alternatives:

• Kimi-K2.5
• GLM-5.1
• MiniMax-M2.7 

Kimi-K2.5 and GLM are open-source, while MiniMax is not. The goal of this article is to highlight how you can run OpenClaw with a lot of different models, highlight how to do it, and how to make your OpenClaw assistant effective.

Why use OpenClaw with open-source models

The main reason I switched from Claude Code to open-source alternatives was simply cost. Anthropic has now blocked subscription tier usage for OpenClaw, and you can now only use Claude with OpenClaw through an API. The API pricing for Claude Opus 4.6 is currently , which quickly racks up the cost.

I thus started looking for alternative solutions that were cheaper and still provided good performance. I first attempted to use OpenAI, which has a 200 USD subscription tier that you can use with OpenClaw. However, I experienced the LLM being quite lazy, and unwilling to resolve problems independently. A lot of times, I had to help the model a lot when solving new problems, which is obviously not ideal when you’re working with an assistant.

If you do a quick Google search for the best OpenClaw models online now, you’ll probably get a list with Claude Opus, followed by some Chinese models such as Kimi-K2.5. These models are a lot cheaper than Claude Opus 4.6, with Kimi-K2.5 priced at 0.6/3 USD per million tokens, around 1/10th the price of Claude Opus 4.6.

Thus, I decided to try out Kimi-K2.5 to see if it worked well and if I could make it an effective OpenClaw assistant.

How to use Kimi-K2.5 in OpenClaw

I started using Kimi-K2.5 in OpenClaw, and it was pretty easy to set up. First of all, I needed access to the Kimi-K2.5 model. You can do this through the official Kimi-K2.5 website. However, I decided to do this through OpenRouter because it provides me with some added flexibility and uptime. When you access Kimi-K2.5 through OpenRouter, you pay around a 10% upcharge because of the middleman cut. However, in exchange, you get easy access to many models, including other Chinese alternatives, and can super easily switch between them.

To set up Kimi-K2.5 in my OpenClaw, I simply fetched an API key from OpenRouter, provided it to my Claude Code instance, and asked it to set up my OpenClaw model to use Kimi-K2.5 instead of the Anthropic models.

One important thing I noticed when switching from using the Anthropic subscription was that you need to remove all references to Anthropic. I.e., I had a previous or existing OpenClaw assistant that was running on Claude Opus 4.6. When I then switched to using Kimi-K2.5, I still experienced OAuth issues even though Kimi-K2.5 was the main model used for my assistant. The reason I found out was because of some Anthropic references, including an Anthropic key in my environment variables. Make sure to remove all of these, so you don’t experience OAuth issues.

Make sure to remove all previous model references, for example to Anthropic, when setting up a new LLM for your OpenClaw assistant

After that, it was pretty straightforward. was able to one-shot the implementation.

Kimi-K2.5 Performance

In this section, I will cover the performance of the Kimi-K2.5, especially compared to Claude Opus and OpenAI GPT-5.4. If I were to completely ignore cost and simply think about performance, I would put them in the following order:

1. Claude Opus 4.6
2. Kimi-K2.5
3. GPT-5.4

However, the gap between 1 and 2 is way smaller in my opinion than the gap between numbers 2 and 3. Kimi-K2.5 is not far away in performance from Claude Opus when it comes to being useful as an OpenClaw assistant.

I would, however, like to note that I did experience Kimi-K2.5 being quite slow at times, which I believe happened because it used more thinking tokens than should be necessary on easy tasks, and this was a recurrent thing I noticed compared to Claude Opus 4.6. However, I was more easily able to ensure that Kimi-K2.5 kept trying and didn’t give up easily on tasks that it should be able to perform.

Thus, overall, if I were completely to the cost, I would probably choose Claude Opus 4.6. However, when Kimmy comes in at 1 tenth of the price, I believe it’s a really strong competitor and can easily compete in a lot of areas with Claude Opus 4.6.

Techniques to optimize OpenClaw

I also want to cover how to achieve better performance with OpenClaw when using open-source models such as Kimi-K2.5. Of course, you have all the standard tips you should do when using OpenClaw, which include:

• Ensuring the model has specific skills for each task that it performs.
• Giving it all the permissions it needs, such as API keys to different services.
• Setting up cron jobs to ensure the model learns from its previous chats. You could, for example, have a daily cron job reviewing all of today’s chats.

Overall, I followed these tips and general OpenClaw tips that I previously followed when using Claude Opus 4.6 with OpenClaw. I didn’t really experience an area where tips that worked for Claude Opus didn’t work for Kimi-K2.5. And I just think OpenClaw is quite language model-agnostic, as long as you’re using a language model that is very capable both with regard to reasoning and agentic capabilities.

Downsides of Kimi-K2.5

Even though my overall experience with Kimi-K2.5 was very good, I would also like to highlight some downsides of the model when using it for OpenClaw.

The first downside is the speed of replies for simple requests. I did very clearly notice that Kimi-K2.5 was quite a bit slower, even though I asked very simple requests, such as “Do you have access to a specific service?” where it should reply with a simple yes. The model was spending a lot of time thinking before providing such simple responses. However, I do think it’s worth noting that even though the model was slow, the most important factor for me is the quality of the output from the model. And the speed is not as important. So even though the speed is unfortunate, it’s not mission-critical.

Another downside I would like to highlight is GDPR compliance. Of course, if you’re using Chinese models through an API, you will not be compliant with GDPR regulations requiring you to stay in the EU, etc. This makes it so you cannot use the model for any customer data, or any data that is of high importance, and stays secure.

The good part of this is that Kimi-K2.5 and other Chinese models are open source, so in theory, you could host them yourself and thus be compliant with GDPR regulation. Though this, of course, requires you to do a lot more setup yourself, for example, setting up a GPU where you can run the model, hosting it, speed will probably be slower, and so on, so this has its downsides as well.

Conclusion

In this article, I’ve discussed how to run OpenClaw with open source models, where I spent most of my time highlighting my experience using Kimi-K2.5. I highlighted how Anthropic banned the use of third-party services for their subscription tier, which forced me to make a change, trying out alternative LLMs to power my OpenClaw assessment. I tried OpenAI’s GPT 5.4 and experienced the model being a bit lazy, and thus tried out other models and had a very good experience using Kimi-K2.5. Pretty much, I highlighted how to make it perform as best as possible and some downsides of the model. I believe OpenClaw Assistants are incredibly powerful and urge you to try it out yourself, especially now that you can run Assistants for much cheaper using language models such as Kimi-K2.5. In my opinion, performance is definitely still very high, and it’s able to be a valuable Assistant.