Matt Wolfe tries Locally AI, an Apple-device app for running downloaded models without a cloud connection. His March 2026 demonstration uses an iPhone 17 Pro and Qwen 3.5. The result is a practical offline brainstorming tool with visible limits. Wolfe says the video is unsponsored.

Download a model that fits the phone (01:36–03:32)

  • The initial setup screen shows fewer models than the full selector. Wolfe skips that screen, opens model selection and finds Qwen 3.5 there.
  • The app’s recording-time recommendations associate the 4-billion-parameter version with an iPhone 15 Pro or newer. It recommends an iPhone 15 for the 2-billion version and an iPhone 14 for the 800-million version.
  • Parameters are the model’s learned values; larger versions generally need more memory. These are the app’s recommendations in the video, not a tested compatibility table for every phone.
  • Wolfe downloads both the 4-billion and 2-billion versions to compare them. The first download takes about five minutes on his Wi-Fi; offline use begins after the model is installed.

Simple answers and broken reasoning coexist (03:33–06:21)

  • Settings expose custom instructions, temperature controls and conversation-history deletion. A Siri shortcut offers another way to send questions to the app.
  • The 4-billion version correctly counts the letters R in “strawberry.” That successful example is followed by a conspicuous failure on a different prompt.
  • Asked whether to walk or drive to a nearby car wash, the 2-billion version compares travel times and misses the need to bring the car. Wolfe’s follow-up does not produce a sensible correction.
  • Brainstorming works better for his purposes. The smaller version quickly produces video ideas, although some suggestions contain odd wording and unexplained acronyms.

Longer thinking has a phone-sized cost (06:23–08:07)

  • Turning on thinking mode makes the model spend longer preparing its response. Wolfe feels the phone warming during the run.
  • As the conversation grows, scrolling becomes choppy. He observes slower interaction with more conversation context, the earlier text the model must consider.
  • Starting a fresh chat restores faster interaction in the next example. He photographs a drink and asks about it, demonstrating image input rather than establishing reliable nutritional advice.

Airplane mode is the clearest offline test (08:09–09:14)

  • Earlier screen mirroring requires Wi-Fi, so Wolfe switches to the phone’s own screen recording for this test.
  • He turns off Wi-Fi and enables airplane mode, then starts a new chat. The model streams a response to a parenting question without an active internet connection in his demonstration.
  • That sequence is stronger evidence for offline operation than the earlier connected examples. It shows one response working after the model download, not every app feature under every network condition.

Voice requires another download (09:16–11:50)

  • Wolfe restores Wi-Fi for screen mirroring before trying voice mode. Voice components download separately because they were not installed by default.
  • The spoken dinner suggestions are usable but ordinary. This section is a connected demonstration, so it should remain distinct from the earlier airplane-mode test.
  • Wolfe presents local processing as a privacy advantage and calls the app free at recording time. Neither claim should be read as a permanent pricing promise or an independent security assessment.

Caveat: this is one phone and a handful of prompts; there are no measured battery results, speed benchmarks or broad accuracy tests.