Engadget guide highlights cost and privacy benefits of running local AI on iPhone

Engadget has published a comprehensive guide detailing the installation and operation of open-weight AI chatbots directly on iPhones, highlighting a shift away from cloud-dependent services. The publication notes that while local models do not match the sophistication of proprietary systems from major labs, they provide distinct advantages in terms of offline functionality, enhanced privacy, and significant cost savings for power users.

The financial argument for local deployment centres on the elimination of recurring subscription fees. According to the guide, running a local model involves a maximum one-time purchase of $5 for apps such as Private LLM, or is entirely free with alternatives like Locally AI. This stands in sharp contrast to cloud-based alternatives, where users may face monthly costs ranging from $20 for OpenAI’s ChatGPT Plus to $100 for Google’s AI Ultra tier, alongside potential rate limits on free accounts.

Privacy remains a primary driver for adopting local AI, as these applications do not require user logins or data sharing with the entities that trained the underlying models. Engadget points out that developers of local apps state they do not collect usage information, whereas proprietary models often utilise user prompts for training purposes unless users actively opt out. Additionally, local chatbots function without an internet connection, a feature unavailable in standard cloud-based interactions.

However, the guide acknowledges significant functional limitations inherent to on-device processing. Local models possess knowledge cutoffs, with Meta’s Llama 3.2 model limited to data from December 2023, and other referenced models such as GPT-5.5 Instant cut off in August 2024. Furthermore, these local systems lack the robust memory features and real-time web search capabilities found in services like ChatGPT and Claude, which can reference past conversations and augment answers with current online data.

Technical requirements vary based on model complexity and device hardware. Engadget specifies that Meta’s 3-billion parameter Llama 3.2 model requires 1.81GB of storage and is recommended for iPhone 15 Pro devices or newer. In contrast, the 1-billion parameter version requires only 695MB and was successfully tested on an iPhone 12. The publication advises that while larger parameter counts yield better responses, they demand more storage and compute power, resulting in slower performance on older hardware.

For users looking to implement these solutions, Engadget recommends two primary applications: Locally AI, which is free and offers an intuitive onboarding process, and Private LLM, which costs $5. Both apps facilitate the download and management of various open-weight models, allowing users to tailor system prompts and select models based on their specific storage and performance constraints.