I am curious...

[Sophist-UK]

Just started reading about this.

I am conservative about security and tried getting openclaw running in a limited docker environment on my new laptop (6gb Nvidia GPU) but couldn't get it to work. But I am concerned about openclaw security, yet still want to get the benefits of an ai assistant.

I also have a capable and underutilized TrueNAS server without a GPU.

And my background is a lifetime (50+ years) in IT so I believe in architecture and security and I am keen to start using AI to do open source coding and for a household assistant.

I plan to try out both the local version of Caal on my laptop and the groq version of this on my NAS and see how capable it is.

But in the meantime, some questions...

1. Is there any benefit from setting up ollama with Redis to cache LLM calls? I suspect (but cannot demonstrate it) that there are a whole bunch of repeated calls and caching might speed things up.

2. I like the Caal concepts of architecture and predefined workflows to give security, and I am assuming that these workflow relate to limiting how it access other local resources in a secure way. And without having tried it yet I am assuming that for general intelligence that only requires access to the internet, Caal is still able to harness the full power of the llm. Is there a feature comparison anywhere of Caal vs. openclaw? Can it do cron tasks? Can you give it a personality? Can it learn (or is that too dangerous)?

3. How feasible is it to have a hybrid local & remote LLM configuration, using local ai for simple stuff and a remote LLM like Kimi 2.5 for deep thinking?

[Sophist-UK]

Also, is the Caal Ministral LLM quantised? My GPU is 6gb not 8gb, so I would like to run a quantised version?

[cmac86]

1. I have a feature in the roadmap that will use a rules engine to deal with simple predictable requests. ex. What time is it? Can be recognizable through regex and then a response generated and spoken through TTS without ever hitting the LLM.

This would be part of a 3 tier approach. Regex for simple calls, small local model for more complex tool calls and larger model (local or remote) for deeper reasoning. This answers your 3 I believe.

2. CAAL has general knowledge and a web_search tool so yes it can answer general questions. Tricky part with a small model that's being trained on tool calling is trying to also keep its conversational abilities and not turn in to just a tool calling robot. I think 8B model may be limited in this way, needs more testing.

For learning - not sure what you mean. Models "learn" through training data, which is what we are doing with fine tuning caal-ministral. Or they learn with context, which is limited depending on VRAM available and the specific model max. Trick is to inject memory into context when it's needed. I don't know the best way to do this yet.

caal-ministral is 4bit quantized. 5.2G on disk. You can try to run it.