MLX Benchmarks

MLX Benchmarks

Measure device LLM performance

0 ratings
Free

Rating summary

Details

Released
November 27, 2025
Updated
November 11, 2026

Features

MLX Benchmarks screenshot #1 for iPhone
MLX Benchmarks screenshot #2 for iPhone
MLX Benchmarks screenshot #3 for iPhone
MLX Benchmarks screenshot #4 for iPhone
iphone
iphoneDuo
ipad
🖼️Get Icon
Icons↘︎

About

MLX Benchmark measures how fast large language models run on your iPhone and iPad, entirely on device, with Apple's open-source MLX framework. STANDARD BENCHMARK Every device runs the same test: a fixed 512-token prompt, then exactly 128 tokens generated with greedy decoding. One warm-up run and three measured runs, with medians reported. Because the input and settings never change, your results are directly comparable with any other device. COMMUNITY Share your standard runs anonymously and compare your device with everyone else's: • Device ranking: median generation speed, prefill speed or time to first token for each model • Models on a device: how speed changes with model size, next to the memory-bandwidth estimate for that device • Comparison: your run next to the community medians of other devices DETAILED MEASUREMENTS • Generation and prefill speed, in tokens per second • Time to first token, tokenization and model load time • Per-token latency as a timeline and a distribution, with P50, P95, P99 and jitter • Memory: model weights, KV cache and activations, and the peak against the memory iOS lets the app use • Thermal state and Low Power Mode, so throttled runs are easy to spot MODELS Browse a catalog of open models supported by MLX, such as Qwen and Llama in 4-bit. Each model is downloaded on its first run; see what is stored on your device and delete what you no longer need. CUSTOM RUNS Try your own prompt, generation length and number of iterations. Custom results stay on your device. PRIVACY No account and no tracking. Only standard runs are shared, and only while sharing is on: device model and hardware (core counts, memory), the model tested, app and OS version, and the measured metrics. Nothing that identifies you.
Show more

What's New in MLX Benchmarks

1.1.0

November 11, 2026

MLX Benchmark 1.1 is a major update. • Standard benchmark: the same 512-token prompt and 128 generated tokens on every device, so results are directly comparable. • Community: share your runs anonymously and see how your device ranks for each model. • New charts: device ranking, speed by model size, per-token latency, memory usage and side-by-side comparison with other devices. • Models tab: browse the catalog, see what's stored on your device and free up space. • Device details: model, CPU cores and memory. • Redesigned layouts for iPad and iPhone Duo. • Custom runs with your own prompt and settings are still there.

More

Developer apps