Apple AI Updates: October 15, 2026
1. Apple Pretrains a 3B Tri-Modal Masked Diffusion Model Over Text, Image, and Audio
Apple. Apple researchers pretrained what they describe as the first tri-modal masked diffusion model from scratch, a 3B-parameter model trained on 6.4T tokens of text, image-text, and audio-text data that handles text generation, text-to-image, and text-to-speech. The NeurIPS-tagged paper (arXiv 2602.21472) studies multimodal scaling laws, modality mixing ratios, noise schedules, batch-size effects, and inference sampling defaults. Its main technical contribution is an SDE-based reparameterization that decouples the physical batch size, chosen for hardware efficiency, from the logical batch size that governs gradient variance, removing the need to search for an optimal batch size. The authors call it the largest systematic open study of multimodal discrete diffusion to date. Source
2. Language-Selective Neurons in Whisper Detect Code-Switching Without Transcripts
Apple. An Apple paper (arXiv 2610.18726) detects code-switched speech, where speakers alternate languages within an utterance, without running multilingual transcription. The method finds hidden-state dimensions in a pretrained speech encoder that respond selectively to one language using monolingual recordings, aggregates their frame-level activations over an utterance, and tunes a decision rule on a small code-switched set. On CS-FLEURS across 14 language pairs, the Whisper-small encoder approach reached a mean F1 of 0.798, versus 0.776 for a CNN baseline and 0.751 for a HuBERT linear probe trained for language identification. The authors position it as a cheap filter for building more representative multilingual training corpora. Source
3. New Apple TV 4K Brings Siri AI With On-Device and Private Cloud Compute Processing
Apple. Apple announced a fourth-generation Apple TV 4K on the A19 chip ($299, in stores November 13) and brought Siri AI, the next-generation Apple Intelligence assistant, to tvOS in beta on Apple TV 4K second generation and later. Siri AI on the device answers follow-up questions in an overlay during playback, controls smart home scenes, and draws on personal context from an Apple Intelligence-enabled iPhone. Requests run on-device or in Private Cloud Compute, and microphone data from the three far-field microphones is processed in Secure Exclaves isolated from tvOS and apps. Apple says some server-side features have daily usage limits and that expanded access will carry a fee in the future; launch support is English in eight countries and not initially in the EU. Source