Ah, ok. I have been using llama since the beginning and ollama was an ok step. glm4.7flash is better at awk than kimi2.5 and qwen3:30b so that’s something 
3 Likes
And pi does work on phone, so you can have ‘local’ (on your desktop/laptop) glm4.7-flash and serve it over wlan to have agent go wild on your phone, sms sending skill should be easy, why buy mac mini for open claw when you can get same thing natively on sfos
Since it’s your own hw/ai giving it access to contacts/messages is much better sell (still too afraid to actually do it, maybe once jp2 is out)
Isn’t that kinda what Mind2 is (was? will be?) supposed to do?
Well kinda, my lousy 4060 laptop can serve glm4.7-flash at 20-30tok/s (ud-q4_k_xl, does slow down as context grows) with 128k context, any mind2 owners can try the same on their hw? Pretty sure cpu only will not come close.
Edit: and even if mind2 does the same and is just interface between phone and local gpu server, this gives the llm chance to execute commands right on your phone, scary as hell, but if you have one spare could be fun
Edit2: I had pi create two skills for itself based on silica reference documentation, it wrestled with curl and grep for a bit but did create something that looks useful, will paste later, maybe can be useful for others, or might be worth giving that task to a sota model, but with pi running on device a much simpler skill would be to just grep /usr/share/harbour* for actual examples of how each silica component is used in real life and go from there
edit3: sailfish-silica skill.md it cooked
sailfish-webview
whether these cause the llm to use sfos components properly more often is still in the air