Model reference

Ollama name	Repository role	Selection
qwen3.5:2b	General explanation/planning model	Default in routing rules
qwen2.5-coder:3b	Code-oriented model	Coding signals or supported explicit selection

These roles are design choices, not guarantees of correctness. The router does not automatically support every newly downloaded model.

Inspect installed models

In the Ollama host’s shell:

ollama list
ollama show qwen3.5:2b
ollama show qwen2.5-coder:3b
ollama ps

list shows installed models; ps shows models currently loaded. Having both downloaded does not mean both fit in GPU memory simultaneously.

Resource planning

Weights, quantization, context length, and concurrent requests affect memory. The project records choosing the smaller general model for a 4 GB GPU because a larger model spilled to CPU. Treat speed figures in experiment logs as host-specific observations.

Record model identifiers/digests, Ollama version, GPU/RAM, context, prompt size, and cold/warm state when comparing latency. Evaluate answer quality separately from speed.

Generation and reasoning output

The backend sends a raw Ollama generation request with think controlled by OLLAMA_THINK. It defaults to false and includes an empty-response fallback. Confirm behavior with the exact runtime/model rather than assuming every model family returns identical fields.

Next: Choose a model or experimental training.