DeepGrove Releases Maple-Preview, a Small AI Model That Runs on iPhone
Maple-Preview shows a design-time ternary approach that keeps inference speed and memory use low without quantizing an existing model, which DeepGrove argues is the wrong path.
Reporting from 1 source: GIGAZINE.
DeepGrove announced Maple-Preview, a small AI model that runs on iPhone. It is a mixture-of-experts model with 20.2 billion total parameters and 1.49 billion active parameters. DeepGrove says it processes 13 times faster than Bonsai 27B on iPhone. The model is open and available under the MIT License.
DeepGrove has released Maple-Preview, a small AI model built to run on iPhone. The company says the model is 13 times faster than Bonsai 27B on the same device, while using only 7.69 GB of memory when handling a 131,000-token context.
Maple-Preview is a mixture-of-experts model with 20.2 billion total parameters and 1.49 billion active parameters. Instead of quantizing an existing model, DeepGrove designed it to run ternary computations from the start, arguing that lowering precision after training limits both performance and efficiency.
The model is open and distributed under the MIT License. DeepGrove also plans to build a system that tunes the model to individual users based on conversation content.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.