Start with Gemma 3 1B. It is 810 MB, runs on every device Nashi supports, and handles drafting, rewriting and everyday questions well. Move up to a 3B or 4B model if your device shows it as a good fit and the answers feel thin. Nashi checks your memory and tells you which models fit.
Which AI model can my iPhone run?
Bigger models answer better. Smaller models answer faster and fit on more devices. Everything else in this post is detail.
Each model in Nashi has a minimum memory figure. Nashi compares it with the memory iOS reports for your device and puts one of three labels on the model:
- Works great on this device: you have at least 1 GB more than the model needs.
- May be slow on this device: you meet the minimum, with little to spare.
- Too large for this device: you don't meet the minimum. Nashi shows no download button, so you cannot download something that will not run.
One catch: iOS reports a little less memory to apps than the number on the spec sheet. A phone sold with 8 GB reports slightly under 8 GB, and a 6 GB phone slightly under 6 GB. Keep that in mind when you read the list below.
What models are in Nashi?
All models are open weights in GGUF format, quantised to 4 bits (Q4_K_M) so they fit in memory with little loss in quality. Nashi runs them with llama.cpp, with the heavy math on the GPU through Metal. Every one has a 4,096-token context window: roughly 3,000 words of conversation the model can "see" at once.
Needs 2 to 3 GB of memory
- SmolLM2 360M, 270 MB. Quick answers anywhere. Basic chat only.
- Gemma 3 1B (QAT), 810 MB. Marked Recommended. Surprisingly capable for its size.
- Llama 3.2 1B, 810 MB. Also Recommended. A solid all-rounder with a slightly different voice from Gemma.
- SmolLM2 1.7B, 1.1 GB. A step up in quality from the 360M version, still light.
Needs 5 GB of memory
- Llama 3.2 3B, 2.0 GB. Noticeably better reasoning and general knowledge.
- Gemma 3 4B (QAT), 2.5 GB. Marked Recommended, and our pick for 8 GB phones. Strong writing.
- Phi-4 Mini, 2.5 GB. The best of the mid-size group at code and structured reasoning.
Needs 8 GB of memory
- Mistral 7B v0.3, 4.4 GB. Big-model quality.
- Llama 3.1 8B, 4.9 GB. Maximum quality in the catalogue. Slow, and worth it.
In practice: on a 4 GB iPhone, SmolLM2 360M works great and the 1B to 1.7B models may be slow. On a 6 GB iPhone, the 1B to 1.7B class works great and the 3B to 4B class may be slow. On an 8 GB iPhone, everything up to 4B works great.
The 7B and 8B models are a different story. Because an 8 GB phone reports a little under 8 GB, Nashi marks them Too large for this device on 8 GB iPhones. They are there for devices with more memory, such as higher-memory iPads. Nashi still asks you to confirm before you switch to one, since they push even those devices hard.
QAT stands for quantisation-aware training. Google trained those Gemma variants knowing they would be squeezed to 4 bits, so they lose less quality in the process than a model quantised after the fact. If you are choosing between two models of the same size, QAT is the tie-breaker.
Which model should I download first?
One model: one Nashi marks Recommended that also shows Works great on this device. Chat with it for a day. Most people find that a 1B to 4B model handles drafting, summarising, rewriting and everyday questions well, and never download a second one.
Download a larger model when the answers feel thin: when it misses the point of a longer instruction, loses track in a long conversation, or hallucinates facts you need to be right. Download a smaller one when you mostly want speed, or when the phone gets warm during long sessions. Before you switch, it is worth trying the habits in getting good answers from a small model; a better prompt often does more than a bigger download.
Does running AI on an iPhone drain the battery?
While it is generating, yes. A 1B model on a recent iPhone writes faster than you can read. A 4B model is comfortable. The larger the model, the fewer words a second, which is fine for a paragraph and tedious for an essay.
Generation is about the most demanding thing your phone can do, so expect it to warm up and use battery the way a 3D game does. That is the trade for having the model on the device: there is no data centre doing the work for you. When the phone gets hot, Nashi slows generation down rather than let it overheat.
How much storage do Nashi models need?
From 270 MB to 4.9 GB, as listed above. Models live in the app's own storage and are excluded from iCloud backup, so a large model does not eat your backup quota. Delete a model from the Models screen whenever you need the space; your conversations stay, and you can download the model again later.
Is a small model good enough?
Every model here is smaller than the ones behind cloud chatbots, and it shows on hard problems. What you get in exchange is a conversation that exists in exactly one place. For a lot of what people actually type into an assistant, that is the better deal.