Implemented: deterministic routing over model names returned by the local Ollama runtime. It does not benchmark models, detect installed hardware, download models, or prove response quality.
- Smart Model and Smart Model · Balanced: prefer installed Qwen for coding and reasoning requests, Gemma for general chat; choose by stable name ordering within the task's family.
- Smart Model · Fast: prefer the variant with the smallest advertised parameter-count label among installed models in the preferred family.
- Smart Model · Quality preference: prefer the variant with the largest advertised parameter-count label within that family.
- Coding tasks prefer a Qwen model with coder in the name when installed. If the preferred family is absent, fall back only to another installed Qwen/Gemma variant.
- A manual model selection overrides Smart Model and may use any installed model.
Important: Smaller labels are not proven faster on a given CPU/GPU; larger labels are not proven more accurate or safe. Memory, quantization, architecture and workload affect actual performance. Missing size tags sort after known tags.
Use the model chooser in the UI or set the chat model field to smart, smart:fast, smart:balanced or smart:quality. These policy strings work with the existing Chat and Smart Model route preview APIs. Route explanations are included with responses.
Benchmark each candidate with real Qwen/Gemma weights on authorized hardware. Evaluate response quality, latency, resource footprint, memory use, thermal constraints, and privacy-aware routing before claiming hardware-adaptive Smart Model features.