Qwen via Ollama

• Chosen model: qwen2.5:7b (4.7 GB). Other sizes exist if needed: 0.5b / 14b / 32b.
• Endpoint: http://localhost:11434 (standard Ollama API).
• Timeouts: 120 s default for analysis calls; longer (e.g. 300 s) for heavy generation.
• Prefer local over cloud when: offline, data can't leave the machine, or bulk work where API cost matters.
• VRAM contention with other GPU services: see the Vram-GPU-OOM-memory-management skill.

Reference implementation (Marvinous)

• OllamaClient.rs in this folder — copy of the production async client (reqwest, timeout, error handling) from /home/matt/Marvinous/src/llm/client.rs.
• Prompt building: /home/matt/Marvinous/src/llm/prompt.rs
• System prompt (Marvin's personality): /etc/marvinous/system-prompt.txt
• Config (model/endpoint settings): /etc/marvinous/marvinous.toml