Pipette is Liquid AI’s on-device foundation model benchmarking platform, designed to help users compare the best models for real-world device performance. It focuses on key metrics such as speed, latency, memory usage, quality, quantization, runtime, and hardware compatibility. The site’s main goal is to answer the question: “What model runs best on device?”
The platform offers a leaderboard and results pages where users can explore benchmark data by device and model. It also highlights charts showing per-model performance, how performance changes with context length, and the memory cost associated with higher throughput. According to the site, speed and memory are measured directly on-device, while model quality is evaluated on an H100 GPU.
Pipette also invites community participation through a preview program. Users can help collect benchmark data using iOS and Android apps or GitHub-based benchmark clients, with community submission support planned for the future. The platform provides links to documentation, feedback, glossary, and related Liquid AI resources, including its foundation models.
The benchmark results shown are generated with Pipette v1, and the methodology has been independently verified by Artificial Analysis. Overall, Pipette is positioned as a practical tool for comparing foundation models across performance dimensions that matter most for deployment on phones and other devices.
Article timeline
Build this article’s timeline
Extract dated events, key actors, and open questions using the same timeline analysis available in Signal.
Pipette is Liquid AI’s on-device foundation model benchmarking platform, designed to help users compare the best models for real-world device performance.
🤖 AI prefill
📅When
2026
🤖 AI prefill
📍Where
United States and H100
🤖 AI prefill
💡Why
Centers on on-device foundation model benchmarks, foundation models, and benchmark data
🤖 AI prefill
Login to answer and earn XP for each gap you fill.
Key Phrases:on-device foundation model benchmarks100%foundation models97%benchmark data92%model performance90%speed, latency, memory, and quality89%quantization87%inference benchmarks85%throughput84%context length80%on-device AI79%
Comments
Login to view and post comments.