콘텐츠로 이동

asiai recommend

Get engine recommendations for your hardware and use case.

Usage

asiai recommend [options]

Options

Option Description
--model MODEL Model to get recommendations for
--use-case USE_CASE Optimize for: throughput, latency, or efficiency
--community Include community benchmark data in recommendations
--db PATH Path to local benchmark database

Data sources

Recommendations are built from the best available data, in priority order:

  1. Local benchmarks — your own runs on your hardware
  2. Community data — aggregated results from similar chips (with --community)
  3. Heuristics — built-in rules when no benchmark data is available

Confidence levels

Level Criteria
High 5 or more local benchmark runs
Medium 1 to 4 local runs, or community data available
Low Heuristic-based, no benchmark data

Example

asiai recommend --model qwen3.5 --use-case throughput
  Recommendation: qwen3.5 — M4 Pro — throughput

  #   Engine       tok/s    Confidence   Source
  ── ────────── ──────── ──────────── ──────────
  1   lmstudio    72.6     high         local (5 runs)
  2   ollama      30.4     high         local (5 runs)
  3   exo         18.2     medium       community

  Tip: lmstudio is 2.4x faster than ollama for this model.

Notes

  • Run asiai bench first for the most accurate recommendations.
  • Use --community to fill gaps when you haven't benchmarked a specific engine locally.
  • The efficiency use case factors in power consumption (requires --power data from previous benchmarks).