Michael Limberger
Need me? Email mike@limberger.ca
AI
The Models
A player without songs
An AI interface without models is like a music player without songs. Ollama is the library. Let's pull some good ones and say when you would use each.
When you pull a model, Ollama downloads the weights (the "brain") to disk. When you chat, it loads those weights into unified memory. The model stays loaded until you switch away or Ollama unloads it. The constraint is RAM. Bigger is usually smarter and hungrier. Smaller is faster and less capable. That is the trade.
Four specialists
Qwen3 30B-A3B is the star for most people tonight. It is a Mixture of Experts model: about 30.5 billion parameters total, and only about 3.3 billion active per token. You get a lot of the intelligence of a large model with the speed and memory footprint of a smaller one.
ollama pull qwen3:30b-a3bAbout 19GB to download. 32GB RAM recommended. 32K context. Strong at reasoning, code, math, creative writing, and general conversation, with 100+ languages. This is the go-to pick on a 32GB Mac.
Qwen3-Next is the heavyweight. About 80 billion parameters total, roughly 3.9 billion active per token, and a 256K context window. That is enough to hold entire books in one conversation.
ollama pull qwen3-nextAbout 50GB on disk. 64GB RAM minimum. Use it for long documents, complex reasoning, and detailed analysis. Skip it if you do not have the hardware and stick with the 30B.
Mistral Small 3.2 is a 24 billion parameter dense model from Mistral AI. Dense means every parameter is active every time, unlike MoE.
ollama pull mistral-small3.2About 15GB. Comfortable on 32GB. 128K context. Strong instruction following, multilingual, and it can see images. One model for text and vision without switching.
Qwen3-VL is the vision specialist. The 8B version is the sweet spot for most machines.
ollama pull qwen3-vl
About 6GB for the 8B build.
Comfortable on 16GB.
256K context.
OCR in 32 languages, spatial reasoning, charts, screenshots, and even video frames if you extract them first.
Lighter option: ollama pull qwen3-vl:4b (about 3.3GB), which runs on practically anything.
Pull what fits, then the hospital metaphor
On 16GB: qwen3-vl:4b and/or qwen3-vl.
On 32GB: qwen3:30b-a3b, mistral-small3.2, and qwen3-vl.
On 64GB+: add qwen3-next.
Downloads take a while, so start them now.
A dense model uses every parameter for every token. An MoE model has expert sub-networks. For each token, a router picks which experts are relevant and the rest sleep. Qwen3 30B-A3B has 128 experts and activates 8 per token. You get the knowledge of roughly 30 billion parameters at the cost of about 3.3 billion. That is why it runs on 32GB while competing with models that need 64GB. Think of a hospital: you do not need every specialist in every room. The ER triages. MoE does that with neural experts.
Run ollama list to see sizes on disk.
Refresh OpenWebUI and they appear in the dropdown.
Next: the interface.