iPhone & iPad
作者： Jakub Rusinowski · 最后更新： 2026年9月1日
AI 教育者、本地 LLM 部署工作坊讲师
The app disappears to the home screen with no dialog, no crash report, and no error.
On relaunch, some apps report that they have loaded a smaller built-in model instead.| 如果你看到 | 原因是 | 前往 |
|---|---|---|
| The app vanishes with no dialog at all | iOS terminated it for exceeding its memory limit — a jetsam kill, not a crash | Fix 2: choose a model one tier down |
| The app reopens having quietly loaded a smaller model | It detected the failed load and fell back deliberately | Fix 2 — the app is telling you the model is too big |
| It loaded fine and died later, mid-conversation | The context grew and the KV cache pushed it over the limit | Fix 4: shorten the conversation |
| The same model runs on your Mac with the same nominal RAM | An iOS app gets a fraction of physical memory, not all of it | Fix 1: read the device table |
| Storage is nearly full as well | A full device makes memory pressure and loading worse | Fix 5, then the storage guide |
There is no crash dialog and no error to search for, which is why this feels so opaque. The app is simply gone and you are looking at your home screen. That absence is itself the diagnosis: iOS did not let the app fail, it removed it.
iOS gives each app a memory limit well below the device's physical RAM, and enforces it with a mechanism called jetsam: exceed the limit and the system terminates the process immediately. There is no swap on iOS. On a Mac an oversized model gets slow because macOS pages to SSD; on an iPhone there is nowhere to page to, so the app is killed instead. That is why there is no dialog — the app was not given the chance to report anything. It is also why a model that runs on a Mac with the same nominal memory will not load on a phone: the Mac can use nearly all of its RAM for one process, and an iOS app gets a fraction of the device's. Some apps notice the failed load on relaunch and quietly load a smaller bundled model rather than failing again, which is a sensible design and easy to miss.
Model size on a phone is set by RAM, not by storage or by chip generation. The tiers below are drawn from published app compatibility guidance — Private LLM's documentation is a usable public reference — and you should check them against the specific app you are using, since bundled quantisation differs between apps. iPhone SE (2nd gen), 3 GB — roughly 1B to 1.5B class only, such as Llama 3.2 1B or Qwen 2.5 0.5B/1.5B. iPhone 12 to 14, 4 GB — up to about 3B. iPhone 15 Pro and newer, 8 GB — up to about 7B–8B, such as Llama 3.1 8B or Qwen 2.5 7B. iPad Pro, 16 GB — up to about 14B, such as Qwen 2.5 14B or Gemma 2 9B. Anything before the iPhone 12 — not recommended for this.
The fastest diagnostic is to load a model one clear step smaller. If a 3B quits instantly and a 1B runs, you have confirmed a memory limit rather than a corrupt download or a bad install, and you can stop investigating everything else. Do this before deleting and re-downloading anything — a re-download of the same too-large model changes nothing and costs you several gigabytes of transfer.
Usually a cliché, and on iOS a real fix for this specific problem. Memory becomes fragmented over days of use, and a large contiguous allocation can fail even when the total free figure looks adequate. A restart clears that. It also does something force-quitting cannot: iOS manages background apps aggressively and unpredictably, so "I closed everything from the app switcher" often is not true in the sense that matters — swiping an app away does not guarantee its memory was reclaimed.
If it loads and dies later, the KV cache is the thing that grew. These apps typically run around an 8K context on iPhone and iPad against 32K on a Mac, and a long conversation approaches that ceiling from inside a memory budget that was already tight. Start a fresh conversation rather than continuing an old one, and where the app exposes a context setting, leave it at the default rather than raising it.
Check Settings → General → iPhone Storage. iOS needs free space to work with, and a device close to full degrades memory management as well as refusing new downloads. Freeing several gigabytes is worth doing before you conclude the model is too big, particularly on a 64 or 128 GB phone carrying photos and a couple of models.
The honest framing, because it saves people a lot of time: an iPhone is not a small Mac. A 3B model on a phone is a genuinely useful tool — drafting, summarising, questions, all offline and private. An 8B on a 4 GB phone is not going to happen, and no setting, no jailbreak and no quantisation trick changes that; the memory limit is enforced by the operating system. If you need a larger model on the move, running it on a machine at home and reaching it over a tunnel is the setup that actually works.
If the model will not download in the first place, or downloads keep resetting, that is a different problem with its own page. If you are still choosing an app or a first model, the phone setup guide is the place to start rather than here.
上报问题时请附上这些信息
Because iOS terminated it rather than letting it fail. Exceeding the per-process memory limit triggers a jetsam kill, which gives the app no opportunity to display anything. The silent disappearance is the diagnosis, not a missing error.
Because an iOS app is allocated a fraction of the device's physical RAM, while a Mac process can use nearly all of the machine's. There is also no swap on iOS, so there is no slow path — it is either within the limit or terminated.
By RAM: 3 GB devices such as the iPhone SE 2nd gen handle roughly 1B–1.5B; 4 GB devices like the iPhone 12–14 reach about 3B; 8 GB devices from the iPhone 15 Pro onward reach 7B–8B; a 16 GB iPad Pro reaches about 14B.
The KV cache grows with the conversation and eventually pushes the process over its limit. These apps typically run around 8K context on iOS against 32K on a Mac, so long conversations approach the ceiling sooner. Start a fresh conversation.
Less than people expect. iOS manages background apps on its own terms, and swiping them away does not guarantee their memory is reclaimed. A full device restart is more reliable, and it also clears the fragmentation that blocks large contiguous allocations.