Which AI models actually run on your iPhone

Most lists of "small models for phones" are copied from a Hugging Face trending page and never tested on hardware. This one is the shipping catalogue from Pocket AI: 34 open models, each with the exact GGUF we download, and the RAM tier the app enforces before it will let you install it.

The Needs column is not marketing. It is the smallest device RAM at which the model passes the memory gate described in How much RAM does an LLM really need on an iPhone?. Below that number the app blocks the download, because the alternative is a phone that restarts.

Quick answer by phone

Apple does not publish iPhone RAM, so those figures are the reported teardown numbers. The app reads your device's real value at runtime rather than guessing from the model name.

The full catalogue

Sorted by the RAM tier that can hold them, smallest first.

Everything is Q4_K_M except Gemma 3 270M, which ships Q8_0 — at 270M parameters the Q4 saving is about 37 MB while the quality loss is very much not negligible. See GGUF quantization explained for why that trade flips at small sizes.

What to actually pick

If you have 4 GB. You have one real choice and it is a 1B model. Llama 3.2 1B is the safe default — 128k context, and Meta's instruction tuning holds up better at 1B than most. Qwen3 1.7B is smarter but leaves you almost no headroom.

If you have 6 GB. This is the sweet spot for 3B models. Llama 3.2 3B for general chat with a huge context window. Qwen2.5 3B if you want coding and maths. SmolLM3 3B if you want visible reasoning traces at a size that still runs fast.

If you have 8 GB. Now the 4B tier opens up and the quality jump is real. Gemma 3 4B is the strongest all-rounder here — good reasoning, 128k context. Phi-4 Mini is noticeably better at maths and code than its size suggests. Qwen3.5 4B is the newest architecture with hybrid thinking.

If you have 12 GB or more. Ministral 8B and Cogito v1 8B are genuinely capable assistants rather than toys. Expect slower tokens per second — you are trading speed for quality, and on a phone that trade is felt.

Things the size column does not tell you

Context length costs RAM too. The KV cache grows linearly with context. A model advertising 128k context will not give you 128k on a phone; the engine clamps it, and the memory budget assumes the clamped value. Treat the context column as a ceiling on the architecture, not a promise about your device.

Parameter count is not quality. LFM2.5 350M at 219 MB is more useful than SmolLM2 360M at 269 MB, despite being smaller, because it is a newer architecture trained for edge inference. Generation matters more than size within the same tier.

"E2B" is not 2B. Gemma 3n E2B and Gemma 4 E2B are effective 2B — they use a larger parameter set with per-layer embeddings that stream selectively. Gemma 4 E2B downloads at 3.2 GB, not the ~1.5 GB a real 2B at Q4 would be. Size the download, not the name.

Thinking models spend tokens before answering. Qwen3, Qwen3.5, DeepSeek R1 and SmolLM3 emit a reasoning trace first. On a phone at 10–20 tokens/second that is a real wait before the first useful word appears. Worth it for hard questions, irritating for "what's a synonym for fast".

Where these come from

Every entry is a specific GGUF file from a specific repo — mostly bartowski and lmstudio-community. Sizes are read back from the Hugging Face API rather than estimated, and each download URL is fetched to confirm it resolves before it ships. If a number here disagrees with what your phone downloads, the number here is wrong and we want to know.