#GoodMorrows: on Wednesday, 11 November, 2026. Who'd have guessed I'd live to see the year almost turning on it's back. In february I thought I was going down for sure. Anyway, seems no work today, so will just fix the things I coded with Gemini ages ago, that I think Claude-code or Sozhami-Openclaw agent could do better. I've yet to test the latter, for the simple reason that I only have 20 dollars on it, and also not the coding agent, because it'd eat through tokens too fast for my liking. I might upgrade if I get more money in November than I did in October, but that remains to be seen, so for the moment I'll just use Claude-Code to fix stuff Gemini couldn't get right. It's cheaper by 15 rand than both Gemini and ChatGPT's pro packages and I can do more with it per session. Okay it has a 5 hour limit, but so has the free claude-client as well. I'm not the fastest of typers, nor really the most level-headed of systems either. I may think one thing one moment and the next decide that that is the wrong approach and think completely different. The most recent example of which is Sozhami-OpenClaw Agent which I wanted on my MiniPC but I wanted it offline/isolated, so used Ollama and Qwen3:3b, which is all good and well, but the AI had so much machinery behind it with qwen that over lunch on Monday I decided I have had enough of waiting 15 minutes for an answer from an AI that I told to do something, so at that moment bought a Claude API-key and switched from high-memory and processor usage to online useage where I can get an answer in 6 to 8 seconds. Yes, that is going to be extra on the pocket, but so is the electricity bill going to be if the pc takes 15 minutes to process things on about 89 percent usage, and full CPU usage. Those things pull power like hungry dogs, so whether I'll pay more on power or more on faster agentic AI, more will be paid eitherway. Catch22. Damned if I do, damned if I don't. Well I could say that Chirette and Ethia is enough, but they're designed for different purposes. The whole point in today's rambling is that whether you use offline or online AI, you're going to be driving your costs up, either with electricity or with paying for your API key not to expire. There was another thing I wanted to squektalk about, but I forgot what it was, because I heard the Outlook ding for email so I guess work has arrived after all.
@mhussain Decibel is a South-African built mini pc. It runs Windows11 Pro, and has the following specs: Device name J-Decibel
Processor AMD Ryzen 7 8745HS w/ Radeon 780M Graphics (3.80 GHz)
Installed RAM 16,0 GB (15,3 GB usable)
Device ID 5B5878F9-314E-4CA8-8CBA-A3A6944F6AF5
Product ID 00331-20355-39275-AA174
System type 64-bit operating system, x64-based processor
Pen and touch No pen or touch input is available for this display
@Sozhami write 16gb 15gb for use.
your best model i would advise you go for is a 7b model
you've got the ram for it.
or gemma 4e4b.
remember the context length the larger the more ram overhead will be used
but the more that will be remembered during your chat
@mhussain How to tell ollama to pull Gemma, or is gemma non-ollama related, more an agent of its own like openclaw, but even openclaw needs ollama to do things.
@Sozhami
wow, cpu inference?
hmm, can't you get your gpu to take on the load of processing?
are you using ollama?
you don't want to use a 3b model for this, many errors would be had
if you tell me the specks of your machine, i could roughly tell you what model to go for
also
ollama it's default context length is 4000 that's 4k
you can change this in the ollama settings.
in your system tray/ notofication area, find the ollama item shift plus f10 it, and press the settings menu item.
here are some items settings wize you may find of interest
Auto-download updates
Automatically download updates when available.
Auto-download updates
Expose Ollama to the network
Allow other devices or services to access Ollama.
the important bit belo
just a warning!, warning!, warning!
the larger the context length you select the more ram used.
Context length
Context length determines how much of your conversation local LLMs can remember and use to generate responses.
4k
8k
16k
32k
64k
128k
256k
also if you want to connect to this ollama instance over your network say from another machine, make sure that setting is enabled.