The best way to Run OpenClaw with Open-Supply Fashions

I began off testing out OpenAI’s GPT-5.4, however encountered challenges with the mannequin being lazy. I’d, for instance, ask it to carry out a job I do know it’s capable of do, and skilled the mannequin merely giving up after just a few makes an attempt.

That is, after all, unacceptable in the case of a useful assistant, so I made a decision to begin making an attempt out different options, and got here throughout a set of Chinese language options:

• Kimi-K2.5
• GLM-5.1
• MiniMax-M2.7 

Kimi-K2.5 and GLM are open-source, whereas MiniMax shouldn’t be. The purpose of this text is to focus on how one can run OpenClaw with quite a lot of totally different fashions, spotlight methods to do it, and methods to make your OpenClaw assistant efficient.

Why use OpenClaw with open-source fashions

The primary cause I switched from Claude Code to open-source options was merely price. Anthropic has now blocked subscription tier utilization for OpenClaw, and now you can solely use Claude with OpenClaw by an API. The API pricing for Claude Opus 4.6 is at the moment , which shortly racks up the price.

I thus began searching for various options that had been cheaper and nonetheless supplied good efficiency. I first tried to make use of OpenAI, which has a 200 USD subscription tier that you should utilize with OpenClaw. Nonetheless, I skilled the LLM being fairly lazy, and unwilling to resolve issues independently. Quite a lot of occasions, I had to assist the mannequin rather a lot when fixing new issues, which is clearly not supreme if you’re working with an assistant.

In the event you do a fast Google seek for one of the best OpenClaw fashions on-line now, you’ll most likely get a listing with Claude Opus, adopted by some Chinese language fashions similar to Kimi-K2.5. These fashions are rather a lot cheaper than Claude Opus 4.6, with Kimi-K2.5 priced at 0.6/3 USD per million tokens, round 1/tenth the worth of Claude Opus 4.6.

Thus, I made a decision to check out Kimi-K2.5 to see if it labored effectively and if I might make it an efficient OpenClaw assistant.

The best way to use Kimi-K2.5 in OpenClaw

I began utilizing Kimi-K2.5 in OpenClaw, and it was fairly simple to arrange. Initially, I wanted entry to the Kimi-K2.5 mannequin. You are able to do this by the official Kimi-K2.5 web site. Nonetheless, I made a decision to do that by OpenRouter as a result of it offers me with some added flexibility and uptime. Whenever you entry Kimi-K2.5 by OpenRouter, you pay round a ten% upcharge due to the intermediary reduce. Nonetheless, in trade, you get quick access to many fashions, together with different Chinese language options, and may tremendous simply swap between them.

To arrange Kimi-K2.5 in my OpenClaw, I merely fetched an API key from OpenRouter, supplied it to my Claude Code occasion, and requested it to arrange my OpenClaw mannequin to make use of Kimi-K2.5 as an alternative of the Anthropic fashions.

One necessary factor I seen when switching from utilizing the Anthropic subscription was that you have to take away all references to Anthropic. I.e., I had a earlier or present OpenClaw assistant that was operating on Claude Opus 4.6. Once I then switched to utilizing Kimi-K2.5, I nonetheless skilled OAuth points although Kimi-K2.5 was the primary mannequin used for my assistant. The explanation I came upon was due to some Anthropic references, together with an Anthropic key in my setting variables. Make certain to take away all of those, so that you don’t expertise OAuth points.

Make certain to take away all earlier mannequin references, for instance to Anthropic, when establishing a brand new LLM to your OpenClaw assistant

After that, it was fairly simple. was capable of one-shot the implementation.

Kimi-K2.5 Efficiency

On this part, I’ll cowl the efficiency of the Kimi-K2.5, particularly in comparison with Claude Opus and OpenAI GPT-5.4. If I had been to fully ignore price and easily take into consideration efficiency, I’d put them within the following order:

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

Nonetheless, the hole between 1 and a couple of is manner smaller for my part than the hole between numbers 2 and three. Kimi-K2.5 shouldn’t be far-off in efficiency from Claude Opus in the case of being helpful as an OpenClaw assistant.

I’d, nonetheless, like to notice that I did expertise Kimi-K2.5 being fairly sluggish at occasions, which I consider occurred as a result of it used extra pondering tokens than needs to be obligatory on simple duties, and this was a recurrent factor I seen in comparison with Claude Opus 4.6. Nonetheless, I used to be extra simply ready to make sure that Kimi-K2.5 saved making an attempt and didn’t quit simply on duties that it ought to be capable of carry out.

Thus, general, if I had been fully to the price, I’d most likely select Claude Opus 4.6. Nonetheless, when Kimmy is available in at 1 tenth of the worth, I consider it’s a very sturdy competitor and may simply compete in quite a lot of areas with Claude Opus 4.6.

Strategies to optimize OpenClaw

I additionally need to cowl methods to obtain higher efficiency with OpenClaw when utilizing open-source fashions similar to Kimi-K2.5. In fact, you will have all the usual suggestions it’s best to do when utilizing OpenClaw, which embrace:

• Guaranteeing the mannequin has particular abilities for every job that it performs.
• Giving it all of the permissions it wants, similar to API keys to totally different providers.
• Organising cron jobs to make sure the mannequin learns from its earlier chats. You might, for instance, have a every day cron job reviewing all of at the moment’s chats.

Total, I adopted the following tips and basic OpenClaw suggestions that I beforehand adopted when utilizing Claude Opus 4.6 with OpenClaw. I didn’t actually expertise an space the place suggestions that labored for Claude Opus didn’t work for Kimi-K2.5. And I simply suppose OpenClaw is kind of language model-agnostic, so long as you’re utilizing a language mannequin that could be very succesful each with regard to reasoning and agentic capabilities.

Downsides of Kimi-K2.5

Though my general expertise with Kimi-K2.5 was superb, I’d additionally like to focus on some downsides of the mannequin when utilizing it for OpenClaw.

The primary draw back is the velocity of replies for easy requests. I did very clearly discover that Kimi-K2.5 was fairly a bit slower, although I requested quite simple requests, similar to “Do you will have entry to a selected service?” the place it ought to reply with a easy sure. The mannequin was spending quite a lot of time pondering earlier than offering such easy responses. Nonetheless, I do suppose it’s price noting that although the mannequin was sluggish, crucial issue for me is the standard of the output from the mannequin. And the velocity shouldn’t be as necessary. So although the velocity is unlucky, it’s not mission-critical.

One other draw back I wish to spotlight is GDPR compliance. In fact, for those who’re utilizing Chinese language fashions by an API, you’ll not be compliant with GDPR laws requiring you to remain within the EU, and so forth. This makes it so you can not use the mannequin for any buyer knowledge, or any knowledge that’s of excessive significance, and stays safe.

The nice a part of that is that Kimi-K2.5 and different Chinese language fashions are open supply, so in idea, you possibly can host them your self and thus be compliant with GDPR regulation. Although this, after all, requires you to do much more setup your self, for instance, establishing a GPU the place you possibly can run the mannequin, internet hosting it, velocity will most likely be slower, and so forth, so this has its downsides as effectively.

Conclusion

On this article, I’ve mentioned methods to run OpenClaw with open supply fashions, the place I spent most of my time highlighting my expertise utilizing Kimi-K2.5. I highlighted how Anthropic banned using third-party providers for his or her subscription tier, which pressured me to make a change, making an attempt out various LLMs to energy my OpenClaw evaluation. I attempted OpenAI’s GPT 5.4 and skilled the mannequin being a bit lazy, and thus tried out different fashions and had an excellent expertise utilizing Kimi-K2.5. Just about, I highlighted methods to make it carry out as finest as attainable and a few downsides of the mannequin. I consider OpenClaw Assistants are extremely highly effective and urge you to strive it out your self, particularly now that you could run Assistants for less expensive utilizing language fashions similar to Kimi-K2.5. In my view, efficiency is unquestionably nonetheless very excessive, and it’s capable of be a beneficial Assistant.