APSI Bench vs Pipette by Liquid

This app is aimed at interested users and testers who want to test the performance of the integrated CPUs. For this purpose, it offers a mixed test for single-core and multi-core CPUs, as they also occur in real life. In the single core test, the app calculates ten million random numbers from a total of six different data types (Int, Float, etc.). Such calculations are crucial for game apps. APSI Bench then calculates one million FFT transformations that are common in audio applications, for example. In addition to this, the app also runs a ray tracing algorithm that uses several CPU cores at the same time. You can set how many cores should be used here. The results of these tests are based on a reference system (currently a Macbook Pro 16 inch with an eight-core Intel i9 CPU). That reaches about 100 points. Every user of the app can test the ratio of their Apple product to the MacBook Pro. A decisive extension is the newly built in long-term measurement. Here the app carries out the ray tracing tests with all available CPUs over a long period of time that can be set by the user (up to 30 minutes). The result is then displayed in graphic form. If problems arise with overheating of the CPU cores, this will be visible in the graphic. Finally, the app is also a technology demonstration. The entire interface was designed in SwiftUI 2, with gorgeous animations being used. The app will also be offered as a universal app for macOS.

How fast can your phone run real AI? Pipette is an on-device AI benchmark app. It runs small models on your phone and measures actual performance without cloud processing. See how quickly your device processes prompts and generates text, measured in tokens per second. Many mobile benchmarks estimate AI performance using synthetic tests. Pipette downloads real, openly available AI models and runs them locally with the inference engine built into the app. The result is a benchmark that reflects how an offline AI assistant, local chatbot, or other on-device AI application would actually perform on your phone. What you can do with Pipette • Run AI models on your phone • Benchmark prefill speed, generation speed, latency, and peak memory use • Test popular model families like Liquid, Qwen, Gemma, Llama, Granite, and Mistral • Compare model sizes and quantization levels such as Q4 and Q5 • See results in tokens per second with per-model breakdowns • Use thermal readiness checks for more consistent runs • Export benchmark results as CSV • Optionally contribute results to Liquid AI's public dataset • And more… Fair, repeatable results Device temperature can have a major effect on performance. Pipette includes a thermal readiness check that waits until your phone is cool enough before starting a benchmark. This helps produce more consistent results across devices and repeated runs. Who Pipette is for • Developers building local AI, offline AI, and on-device AI apps • Hardware enthusiasts comparing smartphones and chipsets • Reviewers who need a repeatable mobile AI benchmark • Researchers and community members who want to contribute to Liquid AI's benchmark database • Anyone curious about how well their phone can run AI locally How it works • Sign in with your email • Choose a benchmark and the context length • Download the AI model • Run the test and view your results AI model downloads may be several gigabytes. A Wi-Fi connection is recommended. Pipette is built by Liquid AI, creators of Liquid Foundation Models, efficient AI models designed for on-device inference.