Most head-to-head model comparisons stitch together numbers from four different model cards, each measured on different hardware with different quantization and different prompts, and call it a benchmark. That tells you nothing about which one to actually run. So this post does the opposite: it compares Llama 3.2 3B, Qwen 2.5 7B, DeepSeek Coder V2 Lite and Gemma 2 2B at a single common quantization, on the specifications each publisher actually documents, and is explicit about which claims are published figures and which are judgement. This is not a generic "best local LLM" roundup, since we already have that 15-model ranking. It is four specific models and a like-for-like comparison of what they cost you to run.
Want the short version? Jump to the summary table. Running one of these on an iPhone? PocketLLM lists different variants: Llama 3.2 1B, Qwen 2.5 0.5B, and Gemma 3 1B. Join the launch list.
Qwen 2.5 7B is the largest general model of the four. For running on your own device, Llama 3.2 3B is the better trade for a phone, at under half the file size. DeepSeek Coder V2 Lite is the coding pick if you have 12-16 GB of RAM. Gemma 2 2B is the lightest of the four, at ~1.6 GB on disk. Pick by your RAM, not by the leaderboard.
PocketLLM is launching soon. Private, on-device AI, starting on iPhone, with more platforms planned. No account required, no in-app tracking, and chat processing stays on your device. Join the launch list and be first in.
How we compared
This comparison uses each model's published specifications and what its publisher optimised it for. It is not a first-party benchmark: we have not run these four models through a controlled suite, and we would rather say so than print numbers nobody can reproduce.
- Footprint: On-disk size at Q4_K_M. A model needs more memory than its file size, because the machine must hold the weights plus a working conversation. The file size is checkable before you download anything.
- Quality: Not compared here. Each model card publishes its own benchmark figures; benchmark scores and everyday usefulness are correlated but not the same thing.
- Intended strength: What each publisher says the model was trained to be good at, which predicts real behaviour better than a single aggregate score.
- Licence: Taken from the model repository, because it decides whether you can use the thing commercially at all.
Generation speed is deliberately absent. It is bandwidth-bound and varies with device, quantization and context length, so a number measured on one machine tells you little about yours. What it does do reliably is scale with model size, which is why a 3B feels markedly quicker than a 7B on identical hardware. When PocketLLM ships we will publish measured figures with the full method attached; see iPhone LLM speed, explained.
The four models, head to head
Llama 3.2 3B (our on-device pick)
Meta's 3B is the all-rounder. It fits in ~2 GB at Q4, plus the memory a conversation needs, which is what makes it the phone-friendly option here. It is our on-device pick because it is one of only two models here that fit on a phone, and the larger of the two. For the deeper architectural comparison against Phi and Gemma, see our Llama 3 vs Phi-3 vs Gemma 2 on iPhone breakdown.
Qwen 2.5 7B (largest general model)
Alibaba's Qwen 2.5 7B is the largest general model of the four. At Q4 it is ~4.5 GB on disk and needs more RAM than that for the weights plus a conversation, so it needs roughly 8 GB to run comfortably: laptop or iPad Pro territory, not a phone. It is also the largest dense model of the four, so expect it to be the slowest on identical hardware. If you have the memory and want the largest general model here, this is it. Setup walkthrough in our run Qwen locally guide.
DeepSeek Coder V2 Lite 16B (code pick)
A mixture-of-experts model with only ~2.4B parameters active per token, so it runs at roughly 3B speeds while holding 16B-class code quality. It is the only dedicated code model of the four. The catch is footprint: ~10 GB at Q4 and 12-16 GB RAM to load, which rules out phones entirely. If code is your workload and you have the RAM, it is the pick here. Full setup in our run DeepSeek locally guide.
Gemma 2 2B (lightest pick)
Google's Gemma 2 2B is the lightest of the four, at ~1.6 GB Q4 plus the memory a conversation needs. It is a 2B, so expect it to trail the larger models on reasoning and code. If you are on a phone, an older iPad, or any device where every megabyte matters, Gemma is the one to pick.
The summary table
| Model | Size (Q4) |
|---|---|
| Llama 3.2 3B | 2.0 GB |
| Qwen 2.5 7B | 4.5 GB |
| DeepSeek Coder V2 Lite | 10 GB |
| Gemma 2 2B | 1.6 GB |
How to read these numbers
The table covers file size only. File size, plus the memory a conversation needs on top of it, decides whether you can run the model at all. Quality is a separate question, and this page does not compare it: check the benchmark figures on each model card. Qwen is the larger model but loses to Llama for phone use, because a 7B needing roughly 8 GB is a non-starter there while a 3B at ~2 GB is comfortable. If you only ever run on a 16 GB+ machine, file size matters less and Qwen, the largest general model here, moves to the top for general use.
Pick by your hardware
On a phone or older iPad (A-series): Gemma 2 2B for speed, or Llama 3.2 3B for the largest model that still fits. Qwen and DeepSeek do not belong here.
On a MacBook Air or iPad Pro M-series (8 GB): Llama 3.2 3B as the daily driver; Qwen 2.5 7B if you want the largest general model and have the memory for it.
On a 16 GB+ laptop or desktop: Qwen 2.5 7B for general work, DeepSeek Coder V2 Lite when you are writing code. Both are comfortable here.
Frequently asked questions
Which is better for on-device use: Llama, Qwen, DeepSeek, or Gemma?
Qwen 2.5 7B is the largest general model of the four, but Llama 3.2 3B is the better on-device pick for most people because it fits in roughly 2 GB at Q4 and runs on a phone. DeepSeek Coder V2 Lite is the pick if your workload is code, and Gemma 2 2B is the lightest if you are memory-constrained. There is no single winner — the answer depends on your RAM and whether you value speed or peak quality.
Is this a first-party benchmark?
No, and it does not claim to be. The comparison uses each model's published specifications, the quantized file size at a common Q4_K_M setting, and what each publisher states the model was optimised for. Comparing on published specs at a fixed quantization keeps it apples-to-apples; quoting throughput numbers nobody can reproduce would not.
Can all four models run on a phone?
No. Llama 3.2 3B (~2 GB Q4) and Gemma 2 2B (~1.6 GB Q4) run comfortably on a modern phone. Qwen 2.5 7B (~4.5 GB Q4) needs roughly 8 GB of RAM, which is laptop or iPad Pro territory. DeepSeek Coder V2 Lite is a 16B mixture-of-experts model that wants 12-16 GB, so it is desktop-class. On a phone, stick to the 1B-3B models.
Is Qwen really better than Llama on benchmarks?
This page does not compare benchmark scores. Each model card publishes its own figures, on different test setups, so check them directly. Qwen 2.5 7B is the larger model. But Llama 3.2 3B is less than half the file size (~2 GB against ~4.5 GB at Q4) and runs on a phone, which is why it is our on-device pick. Benchmark quality and practical on-device value are different questions.
Which model should I pick for coding?
DeepSeek Coder V2 Lite if you have the RAM. It is the only dedicated code model of the four, and its mixture-of-experts design holds 16B-class code quality while running at roughly 3B speeds. If you are on a phone or a light laptop, Qwen 2.5 Coder 1.5B is the best small coder. Llama 3.2 3B handles light scripting fine but is not a dedicated code model.