Benchmark candidate local models for smdu AI tasks

[ScottMorris]

Goal

Benchmark local model candidates for smdu AI tasks and produce a shortlist.

Questions to answer

• Which model sizes are viable on common developer machines?
• What latency/memory profile is acceptable for interactive TUI usage?
• Which tasks require embeddings vs generation vs classification?

Candidate families

• Qwen2.5 instruct variants (0.5B, 1.5B, 3B)
• Qwen2.5 Coder instruct variants (for path/action reasoning)
• compact embedding models compatible with transformers.js

Evaluation matrix

• cold start time
• warm inference latency
• memory footprint (idle and active)
• quality on task-specific prompts
• download size and first-run impact

Acceptance Criteria

• at least 3 viable candidates benchmarked end-to-end
• one recommended default model per task type (summary/query/tagging)
• explicit fallback strategy for low-memory environments