# localaiprices

> Live prices (USD and EUR), stock, and measured tokens/sec per $1,000 and per €1,000 for local-AI hardware: AMD Strix Halo mini PCs, NVIDIA DGX Spark/GB10 boxes, Apple Mac Studio, and RTX GPUs. Updated on a schedule; all data is free static JSON.

## API
- [All machines with prices and value scores](https://localaiprices.com/api/v1/skus.json): JSON, sorted by tok/s per $1k
- [Latest price observations](https://localaiprices.com/api/v1/prices/latest.json)
- [Benchmark sources](https://localaiprices.com/api/v1/benchmarks.json): llama.cpp community threads plus llm-speed.com (CC BY 4.0)
- [Cloud GPU hourly prices](https://localaiprices.com/api/v1/cloud.json)

## Pages
- [Comparisons](https://localaiprices.com/compare/): 18 head-to-head pages ([JSON](https://localaiprices.com/api/v1/compare.json))
- [Models](https://localaiprices.com/models/): memory, fit and measured tok/s for 16 popular open models ([JSON](https://localaiprices.com/api/v1/models.json), [benchmarks JSON](https://localaiprices.com/api/v1/model_benchmarks.json))
- [Ranking](https://localaiprices.com/)
- [Rent instead](https://localaiprices.com/rent/)
- [Benchmarks](https://localaiprices.com/benchmarks/)

## Notes
- Small-model metric: llama.cpp llama-bench, Llama-2 7B Q4_0, tg128 (text generation tok/s); best of FA on/off as posted
- Large-model metric: gpt-oss-120b MXFP4 (59 GiB weights), text generation tok/s at empty context, single request (Needs about 59 GiB for weights plus context and OS, so devices with 64GB or less are marked 'doesn't fit'.)
- kind=gpu-card items are graphics cards only and need a host PC
- Fields ending in _converted=true were converted at the ECB EUR/USD reference rate (1.1253, 2026-10-16).