On-device AI leaderboard

Inference engines and models, ranked on the phones and laptops people own. Every rank links to the experiments and raw data behind it.

Next Not ranked yet

Which small model is good enough for a task, on the device you have?

Small language models on everyday app tasks: routing, extraction, tool calls. Filter by engine, or let each model run on its best compatible engine.

Each row
Model configurations: a model at a quantization, on an engine. By default each model runs on its best compatible engine for the objective; filter by engine to fix it.
Views
Overall · Desktop · Mobile
Filters
Engine · Hardware · Workload · Quantization
Measures
Task success rate · Time to a complete answer · Peak memory

Rank by

  • BalancedTask success first, then latency and memory.
  • SpeedTime to a complete answer, among models that pass the task.
  • MemoryPeak memory, among models that pass the task.

Its methodology comes first. Nothing is ranked here until its experiments are published with their raw data.