Which open models actually fit on an iPhone, what quantization really costs you, and what each new release changes. Written by the people who ship Pocket AI.
Subscribe by RSSiFLYTEK's Ciyuan Xinghuo open-sourced a 1.7B and a 4B dense model on September 1 under Apache 2.0 — the 1.7B fits down to a 4 GB iPhone, the 4B needs 8 GB, and the advertised 1M-token context is not something either one can actually use on a phone.
Read →Six iPhone apps that run a real LLM on-device with no internet — Enclave, Locally AI, Private LLM, PocketPal, Apollo and Pocket AI — compared on price, app size, iOS requirement and what they actually do offline.
Read →OpenBMB's MiniCPM5-2B shipped September 7 as a plain Llama-architecture checkpoint — no fork, no waiting on llama.cpp support — and it is small enough to sit comfortably on a 6 GB iPhone.
Read →MBZUAI's IFM released a 3.7B and a 7B model built for phones on September 3, but upstream llama.cpp does not support the architecture yet, so nothing built on stock llama.cpp — Pocket AI included — can load them today.
Read →A 2 GB model does not need 2 GB of RAM. Here is the arithmetic we actually enforce in Pocket AI, the incident that produced it, and the number your iPhone can really hold.
Read →A model-by-model reference for on-device LLMs — download size, the RAM each one really needs, and context length. 34 open models, sorted by the phone that can hold them.
Read →What the letters in Q4_K_M actually mean, why almost every phone model uses it, and the one case where we deliberately ship Q8_0 instead.
Read →