Shorter and more technical than the Build Journal — this is where specific observations, gotchas, and "huh, interesting" moments get logged as they happen.

Note 001 — Ollama on AMD Ryzen AI Platform

During installation, Ollama copied AMD ROCm/HIP libraries including:

...Ollama\lib\ollama\rocm_v7_1\ggml-hip.dll

Observation: No separate AMD ROCm install was required — Ollama bundles what it needs for the Ryzen AI platform.

Note 002 — Ollama Launch UI

Ollama displayed a Launch page with agent integrations such as Claude Code, Codex App, Hermes Agent, OpenClaw, and OpenCode.

Ollama Launch page listing agent integrations: Claude Code, Codex App, Hermes Agent, OpenClaw
Ollama's Launch page — one-command agent installs

Didn't expect Ollama to ship its own agent launcher, Claude Code included. One less thing to configure by hand.

Note 003 — Ollama Install Footprint

Ollama consumed about 5 GB of disk before any models were downloaded.

Note 004 — Model Downloads Use Layers

Ollama downloads models in layers, similar to how container images work.

Note 005 — Qwen2.5-Coder Behavior

Qwen2.5-Coder:14B is a coding-oriented model. In Open WebUI, it attempted to emit JSON-style tool calls in response to simple conversational questions. This appears to be expected model behavior, not a broken installation.

Important lesson

Coder models are not necessarily ideal as everyday chatbots — they're tuned for code and tool-use patterns, which can leak into casual conversation.

Note 006 — Need a Conversational Model

Next step: install a general conversational model, likely qwen3:8b or qwen3:14b, for everyday use in Open WebUI — see Models for the plan.