You need about 3 GB of RAM to run a small but useful language model on a phone, 6 GB to run a good 4-billion-parameter model comfortably, and 8 GB or more for 9-billion-parameter models. The model file has to sit in memory alongside the operating system and your other apps, so your phone’s total RAM needs to be well above the file’s size, not merely equal to it.
The working rule is model size plus about 2 GB of headroom. Everything below explains where that comes from, what quantization does to the numbers, and which model to pick once you know your RAM.
The quick answer by phone #
| Your phone’s RAM | Examples | Models that run well |
|---|---|---|
| 3 to 4 GB | iPhone 13, budget Android | Up to about 2 GB of model (Gemma 4 E2B, Ministral 3 3B, Qwen 3.5 4B at a squeeze) |
| 6 GB | iPhone 14 and 15, mid-range Android | 4B models comfortably (Qwen 3.5 4B, Gemma 4 E4B) |
| 8 GB | iPhone 15 Pro, iPhone 16 lineup and later | 9B models (Qwen 3.5 9B, GLM 4.6V Flash) |
| 12 GB or more | Recent Android flagships | 9B models with room to spare |
How do you work out how much memory a model needs? #
A model’s memory use has three parts.
1. The weights #
The weights are the model’s learned numbers, and their size is roughly:
parameters × bits per weight ÷ 8 = bytes
At full 16-bit precision, a 4B model is 4 billion × 16 ÷ 8 = 8 GB. No ordinary phone can hold that alongside everything else. Compressed to about 5 bits per weight, the same model is around 2.5 GB.
2. The context (KV cache) #
While it’s generating, the model keeps a working record of the conversation so far, called the KV cache. It grows with the context length, so a long chat or a pasted document costs memory on top of the weights. That’s why local apps let you set a context size, and why very long chats slow down or start forgetting their beginning.
3. Everything else on the phone #
iOS and Android don’t hand all the RAM to one app. The system, the keyboard, background apps and the AI app’s own interface all need space. If a model leaves too little, the phone closes the app when you switch away, or closes other apps to make room.
Hence the plus-2-GB rule. A 2.7 GB model is comfortable on a 6 GB phone and workable on a 4 GB one.
What is quantization? #
Quantization stores each weight with fewer bits. Instead of a precise 16-bit number, the model uses an 8-bit, 4-bit or even 2-bit approximation, grouped into blocks that share a scale. It’s how a model that would need 8 GB fits in 2.5 GB.
| Level | Approx. bits per weight | 4B model size | Quality |
|---|---|---|---|
| 16-bit (F16) | 16 | ~8 GB | Original |
| Q8 | ~8.5 | ~4.3 GB | Practically identical to the original |
| Q5 / Q4 | ~4.5 to 5.5 | ~2.5 to 2.9 GB | Small loss; the usual choice for phones |
| Q3 / Q2 | ~3 to 4 | ~1.6 to 2 GB | Noticeable loss, unless the model was trained for it |
Real file sizes vary a little, because some layers are kept at higher precision. If you download models yourself, the level is the code in the filename, and GGUF quantization explained decodes them.
The exception in the last row matters. Quantization-aware training teaches a model during training to cope with low precision, so it loses much less quality when squeezed hard. Google’s Gemma 4 builds for phones are trained this way.
How are small models made in the first place? #
Quantization is only the last step. Model makers also use:
- Distillation. A small “student” model is trained to imitate a much larger “teacher.” Most good small models owe a lot to this; see what is a distilled AI model.
- Pruning. Removing weights or layers that contribute little.
- Efficient architectures. Gemma 4’s per-layer embeddings, for instance, mean the parameters doing work per token (about 2.3B for E2B and 4.5B for E4B) are fewer than the total stored in the file, which is what the “E” refers to.
A 4B model in 2026 is far more capable than a 4B model was in 2023. Within one family, though, size is still the best single predictor of quality.
One thing that doesn’t help on a phone: mixture-of-experts. An MoE model only uses a slice of its weights per token, which makes it fast, but every weight still has to be in memory. A 30B MoE model doesn’t fit a 8 GB phone just because 3B are active at a time.
Real sizes: the current phone models #
These are the downloads and minimum RAM for the models in the Personal LLM catalog. All of them can read images once you add their image-support file.
| Model | Download | Minimum RAM | Comfortable RAM | Best for |
|---|---|---|---|---|
| Qwen 3.5 0.8B | 0.81 GB | 2 GB | 4 GB | Very low-end phones; fast, limited |
| Gemma 4 E2B | 2.04 GB | 3 GB | 5 GB | Everyday chat, 140+ languages |
| Ministral 3 3B | 2.15 GB | 3 GB | 5 GB | Documents and screenshots, long context |
| Qwen 3.5 4B | 2.74 GB | 3 GB | 5 GB | The best first download for most phones |
| Gemma 4 E4B | 3.00 GB | 5 GB | 7 GB | Harder questions, hybrid thinking |
| Qwen 3.5 9B | 5.68 GB | 6 GB | 8 GB | Best answers in the catalog, slower |
| GLM 4.6V Flash | 6.17 GB | 8 GB | 10 GB | Strong vision on high-end phones |
“Comfortable” is the minimum plus 2 GB. Adding image support to a model takes another 195 MB to 1 GB of storage and a little more memory while it’s in use. For the storage side of the question, see how much storage AI models need.
How much quality do you lose with a smaller model? #
It depends entirely on the task.
| Task | Small (under 1B) | Mid (2B to 4B) | Large (9B) |
|---|---|---|---|
| Rewrite or shorten a message | OK | Good | Good |
| Summarize a page of text | Patchy | Good | Very good |
| General knowledge questions | Often wrong | Mixed | More reliable |
| Multi-step math or logic | Weak | OK with thinking mode | Good with thinking mode |
| Write or explain code | Weak | OK for short snippets | Good |
| Read a dense document or chart from a photo | Weak | OK | Good |
Smaller models make more factual mistakes, lose the thread in long instructions, and know less about niche topics. For rewriting, short summaries and quick questions, a 4B-class model is often indistinguishable from a 9B in daily use. Thinking mode narrows the gap on reasoning problems at the cost of waiting longer.
Where every phone-sized model is weak is recall. Treat a local model as a capable assistant for text you give it rather than an encyclopedia, and paste in or attach the facts you want it to work from.
What happens if the model is too big? #
- It won’t load, or the app warns you before you try.
- It loads but crawls, because the phone is juggling memory.
- The app gets closed when you switch away, then reloads slowly when you return.
- Other apps get closed to make room, so your browser tabs reload.
If you see any of these, drop to the next size down. A smaller model that runs smoothly is more useful than a bigger one that stutters. Local AI app crashing or model won’t load covers the diagnosis.
Does more RAM make it faster? #
Not directly. RAM decides what fits. Speed depends mostly on your chip’s GPU and memory bandwidth, plus the model’s size: a 9B model generates more slowly than a 4B on the same phone, because each token means reading nearly the whole file from memory. Once a model fits with headroom, extra RAM mainly means you can keep other apps open. Local LLM speed on your phone has the measurement side.
How do you find your phone’s RAM? #
- iPhone: Apple doesn’t list it in Settings. Roughly: iPhone 13 has 4 GB, iPhone 14 and 15 have 6 GB, and iPhone 15 Pro and the iPhone 16 lineup onward have 8 GB or more.
- Android: Settings → About phone, or on Samsung, Settings → Device care → Memory.
Or skip the lookup. Personal LLM reads your phone’s actual RAM and badges every model in its catalog: “Fits your device” when you have at least 2 GB of headroom over the minimum, “Should run” when you merely meet it, and a plain “Needs N GB RAM” when you don’t. It checks free storage before a download starts, and the chat header shows how much of the context window you’ve used. The step-by-step setup is in our guides for iPhone and Android.
If you’re buying rather than checking, the best phones for running a local LLM ranks current options by RAM and GPU support.
Frequently asked questions #
Can I run an LLM on a phone with 4 GB of RAM? #
Yes. Models around 2 GB, such as Gemma 4 E2B or Ministral 3 3B, run on 4 GB phones, and a sub-1B model runs on less. Close other apps first, and expect simpler answers than larger models give.
How much RAM do I need for a 7B or 9B model on a phone? #
Plan on 8 GB of physical RAM. A 9B model at typical phone compression is 5.5 to 6 GB and needs headroom for the system and the conversation’s working memory. Some builds start on 6 GB phones but run tight.
Does quantization make the AI dumber? #
A little. 8-bit versions are nearly indistinguishable from the original, and 4-bit versions lose a small amount of quality at roughly a quarter of the size. Below 4 bits the loss is noticeable, except for models trained specifically for low precision.
Does virtual RAM help run AI models? #
No. “Virtual RAM” or “RAM Plus” on Android uses storage as overflow memory, which is orders of magnitude too slow for a model’s weights. Only physical RAM counts.
What’s the best first model to download? #
For most phones with 3 GB or more, a 4B model such as Qwen 3.5 4B: fast, capable and able to read images. If your phone has 8 GB or more, add a 9B for the harder questions and switch between them.