I have set up openclaw on my slightly older server that has no GPU, but 64GiB RAM. I tried with various different models, and I think I now found something that works ok.
The problem is of course that most models are censored, and I can only run small models of which most don’t work that well with openclaw. I have arrived at ollama/huihui_ai/Qwen3.6-abliterated:35b, which does work quite well. It’s a mixture of experts model, so responds fast even though it is quite big. Responses to simple prompts come after a minute, more complex maybe take ten, or sometimes even more. But for me that’s fine, given that I’m just using unused computation.
One issue with this model is of course that it is Chinese, and it does for example avoid the topic of the Tiananmen Square Massacre. But I’m not intending to talk about this kind of heavy stuff with it, so it’s not that bad. And when asked directly, it also searches the web for information about it.
For roleplaying it works quite well so far. I am still testing out the tool calling capabilities.
Before I tried ollama/huihui_ai/gemma-4-abliterated:26b, but while it talked nicely, it was bad at calling tools and updating e.g. the SOUL.md and IDENTITY.md file, and also tried to write information about me into the IDENTITY.md file instead of the USER.md file. So I looked for alternatives.
But I am interested in your setups. What models work well for you, and what kind of hardware do you use for them? Do you have any experience to share, or any tips you can give?
I don’t have experience nor do I want to get into it, but openclaw is a very token heavy application and you can probably do a lot better performance wise.
The reason it takes so long for first message is the amount of prompt openclaw requires. It’s almost certainly overkill, and with low prompt processing it will spend most of its time reading memory and soul files and config info for all of its tools.
I think you could just install earandil pi, set a prompt, and add a memory plugin and this will be faster. This is what openclaw uses if you strip back all the extras.
Also look for mixture of experts, they’re far less GPU intensive, so you’ll get better mileage. No idea what models are out there, but like even an a3b or a4b will be slow but maybe usable.
Again, I do not want to get into what this use case lacks or doesn’t work for, that’s your business and I don’t want to get into it. I’m just saying openclaw is probably like driving a 12 ton Humvee to the grocery store.