Foundry
AI on your own machine.
See what every answer costs.
macOS, Windows, Linux. Coming soon
Local
Runs on your laptop. No account, no cloud.
Measured
Joules and tokens per second, under every answer.
Tuned
The settings our research found cheapest.
- Models
- Qwen2.5 Instruct, 1.5B and 7B (Apache-2.0). Our own Trace and Lattice models replace them when they beat them.
- Default
- 4-bit weights, the cheapest per correct answer we have measured (note 20).
- Fast mode
- Speculative decoding with a small draft model: up to 34% less energy per token (note 21).
- Energy
- Read from the chip on Apple Silicon. Where a machine cannot measure it, Foundry says so.
- Engine
- llama.cpp, under a small native app.