AI

[2507.13575] Apple Intelligence Foundation Language Models: Tech Report 2025

Authors:Ethan Lee, Anders Bouson Lindbo Larsen, Chen Chang, Xio Chu, John Chen, Diane Ang Yap, Narindran Raghavan, Shankai Chang, Margate Bulaler, Irai Yildiz, John Bebsip, Hana Gillis Coleman, Matteo Ronchi, Peter Ramadi, Rashik, Anthony, LI, Yuli Yang , Emad Soroush, ZHIYUN Lu, Crystal Xiao, Rong Situ, Jordan Huffaker, David GRIFIFIHS, Zaid Ahmed, Peng Zhang, Daniel Parilla, Asaf Liberman, Jennifer Mallalieu, Persa Mazaheri, Qibin Chen, Manjot Bilkh, Fitzmaurice, Thomas Voice, Jerey Liu, Josh Shafer, ShiWen Zhao, Prasanth Yadla, Farzin Raseeh, Pengsheng Guo, Arsalan Farooq, Jermy SNOW, Stephen Murphy, Tao Lei, Minsik Cho, George Horrell, Sam Dodge, Linds Hishlop, Isaaria Raghavan, Sasha Serofika, Mandana Saybi, Fay Lau, Max Lam, T. Astrauskas, Floris Weers, Josh Gardner, Mira Chiag, Yi Zhang, Pulkit Agrawal, Tony Sun, Quentin Kenebroek, Matthew Hopkins, Bugu Wu, Tao Jia, Chen Chen, Xingyu Zhou, NANZHU Wang, Peng Liu, Jonathan Janc, Ziroy Wang, Chan, Xiaoy Ren, Feng Nan, Josh Eloman, Dong Yin, Youssef Gourene, Jeff Lay, Iran Fay

And others. (297 No additional authors appear)

View the PDF file for the paper entitled Apple Intelligence Foundation Models: Tech Report 2025, written by Ethan Lee and 396 other authors

PDF HTML (experimental) view

a summary:We offer multimedia linguistic models with Apple intelligent via Apple devices and services: IA 3B-Parameter model on the Apple improved device through architectural innovations such as CV-CACHH and training on a 2-bit quantity; The II is a developed server model based on the PT-MOE transformer, the new parallel mix that combines parallel, the branching account of experts, and the intertwined global local interest to provide high quality with a competitive cost on the Apple Special Cloud Calculation platform. Both models are trained in multi -language and multi -media data sets widely by crawling on the responsible web, licensed companies, and high -quality synthetic data, then they are more improved with learning subject to supervision and reinforcement on a new simultaneous platform. The resulting models support several additional languages ​​while understanding images and implementing tool calls. In general standards and human assessments, both the server model and the device model correspond to or bypass open -sized foundation lines.

The new basis framework centered around Swift Generation Generation, inviting restricted tools, and setting a Lora transformer, allows developers to integrate these capabilities with a few lines of software instructions. The latest developments in Apple’s intelligence models are based in our responsible approach to artificial intelligence with guarantees such as filtering content and evaluation of the sites, in addition to our commitment to protecting the privacy of our users through innovations such as Cloud Compute.

The application date

From: Zhifeng Chen [view email]
[v1]

Thursday, 17 July 2025 23:37:19 UTC (966 KB)
[v2]

Tuesday, 26 August 2025 04:02:11 UTC (966 KB)
[v3]

Wed, Aug 27, 2025 16:34:47 UTC (967 KB)

Don’t miss more hot News like this! AI/" target="_blank" rel="noopener">Click here to discover the latest in AI news!

2025-08-28 04:00:00

Related Articles

Back to top button