Diagram of the BeeLink (AI-Lab-B) stack: Open WebUI on top of Ollama, which serves three models — Qwen Coder, DeepSeek, and Gemma.
The current model lineup on Lab PC 2 (AI-Lab-B)
ModelSizeRoleStatus
Qwen2.5-Coder14BCoding-orientedRunning on Node A
DeepSeek-R17BReasoning + web searchRunning on Node A
Qwen38BGeneral conversationalRunning on Node A
"Gwen"—General chatRunning on Node B
Gemma 312BVision-capableRunning on Node A

Qwen2.5-Coder:14B

The first model installed on the lab's first node. A coding-oriented model — strong at writing and reasoning about code, but it noticeably leans toward structured, tool-call-style output even for casual questions in Open WebUI (see Engineering Notebook, Note 005).

Lesson learned

Coder models aren't automatically good everyday chatbots. Different jobs call for differently-tuned models.

Qwen3:8b — the everyday conversational model

qwen3:8b is now the default general-purpose chat model on Node A, sized to run comfortably on the Beelink's 32 GB of RAM. qwen3:14b remains an option if more headroom is ever needed.

See a sample conversation with qwen3:8b — general trivia, a creative-writing test, and the model reasoning through some real Open WebUI tool calls.

Gemma 3:12b — the lab's first vision model

Added August 8, 2026 to Lab PC 2 (Node A), gemma3:12b is the lab's first vision-capable model — it can read an uploaded image, not just text. On a real photograph it was accurate and specific, correctly describing detail down to a chair leg at the edge of the frame. It's set to public so other lab members can select it too.

Diagram showing an uploaded photo entering Open WebUI, passing through Ollama, and routing specifically to Gemma — the only one of the three models on Lab PC 2 that can process images.
An uploaded image routes through Ollama to Gemma specifically — the other two models on this node can't process it

Lesson learned

New models inherit Open WebUI's full Builtin Tools defaults, which can be wrong for a model that doesn't support them. Gemma 3 has no tool-calling support, so its first image upload failed with a "does not support tools" error — that presented as a vision failure but wasn't one. Fixed by disabling the tool-related capabilities and keeping Vision, File Upload, and File Context enabled.

One consequence worth noting: because Gemma 3 can't use tool calling at all, it can't be hooked up to Open WebUI's web search. The lab's vision model and its search-capable model have to stay separate models.

Choosing a model size

Rough guidance for hardware like the Beelink SER9 (32 GB RAM, no discrete GPU):