Which AI Models Run On Your Phone

A practical map of on-device language models: what fits in 4, 6, 8, 12 and 16 GB of RAM, which formats TokForge runs, and where to find speeds measured on real handsets.

The short version: your phone's RAM decides which models will load, and the chip decides how fast they answer. Everything below is about matching those two things to a model that is actually worth using.

Model pages, with measured numbers

Each of these carries the RAM floor the app actually enforces, the download size from the shipping catalogue, and a live table of what real handsets have recorded for that model.

Model	Format	Download	RAM floor
Qwen3.5 0.8B	MNN	0.6 GB	4 GB
Qwen3.5 2B	MNN	1.4 GB	6 GB
Qwen3.5 4B	MNN and GGUF	2.5 to 2.8 GB	8 GB
Gemma 4 E2B	MNN	3.6 GB	8 GB
Qwen3.5 9B	MNN and GGUF	5.2 to 6.8 GB	12 GB

Not sure which of these your handset can take? The phone RAM calculator runs the app's own gate against any model in the catalogue and is explicit about how wide its error bars are.

Start with how much RAM your phone has

This is the constraint that matters most. A model has to fit in memory alongside Android or iOS and everything else running. TokForge reads your RAM on first launch and narrows the catalog for you, but here is the shape of it.

Phone RAM	Sensible model size	Rough download	What it is good for
4 GB	Under 1B	around 0.4 GB	Quick questions, lookups, short rewrites. Do not expect long reasoning.
6 GB	2B to 3B	1 to 2 GB	Everyday chat and writing help. This is where local AI starts feeling useful.
8 GB	around 4B	2 to 3 GB	The everyday sweet spot. A good balance of answer quality and speed.
12 GB	8B to 9B	around 5 GB	Noticeably better answers, with headroom for longer conversations.
16 GB and up	14B	around 10 GB	The best quality that is practical on a phone. Slower, but coherent over long context.

Download sizes vary with the format and the quantization, and TokForge shows the exact figure before you commit to a download. If your phone has 3 GB or less, on-device models will fight the operating system for memory and the experience is poor.

Model families TokForge runs

TokForge is not locked to one vendor's models. It runs standard formats, so the catalog tracks what the open-weights community actually releases.

• Qwen 3 and Qwen 3.5 across the small sizes, from under 1B up through 4B, 8B, 9B and 14B. The small Qwen models are the workhorses of on-device chat right now.
• Gemma 4, including the E2B class built for edge devices.
• Llama 3.1 and 3.2 instruct models, including the 3B that many people start with.
• Mistral and Mistral Nemo for larger-context work on phones with the RAM for it.
• DeepSeek-R1 distills for reasoning-style output at small sizes.
• SmolLM2 and other sub-1B models for 4 GB devices.
• Mixture-of-experts models such as the Qwen3 30B-A3B class, which activate only a fraction of their parameters per token.
• Abliterated variants of the general-purpose models, for fiction and character work where the base model refuses too readily.
• Vision models: Qwen3-VL and Gemma-4 Omni run natively, and SmolVLM ships as a sidecar so text-only models can still answer questions about a picture.
• Image generation: Stable Diffusion 1.5 and LCM variants run on-device, with optional NPU acceleration on supported Snapdragon hardware.

The in-app model hub also browses Hugging Face directly, so a model released this morning is reachable without waiting for us to add it.

GGUF or MNN

TokForge ships two on-device engines and picks between them per device and per model.

• GGUF, run through llama.cpp. The widest ecosystem by far. If you have seen a model on Hugging Face with Q4_K_M in the filename, that is GGUF. It is the right choice when you want a specific community quantization or an unusual fine-tune.
• MNN, Alibaba's mobile inference engine, with OpenCL and Vulkan GPU paths. Strong on some mobile silicon, and the format TokForge's curated bundles usually use.
• Remote, an optional connection to your own OpenAI-compatible server. Off unless you configure it, and it is the only mode that sends anything off the device.

You do not have to choose. TokForge profiles your chipset and routes accordingly. If you want to understand the trade-offs anyway, read Vulkan vs OpenCL vs CPU.

Quantization, and why phone advice differs from desktop advice

Quantization shrinks a model by storing its weights at lower precision. Q4_K_M is the common default, Q5 and Q8 keep more precision and cost more memory, and the IQ family squeezes harder at some quality cost.

Most quantization guides you will find are written for desktop GPUs, where the question is simply whether the model fits in VRAM. A phone is a different problem. Memory pressure is shared with the operating system, the memory bandwidth is far lower, and sustained compute makes the handset hot enough to throttle itself. A larger quantization that technically fits can end up slower and less stable in practice than the smaller one below it. When in doubt on a phone, take the smaller quant of a larger model rather than the larger quant of a smaller one, then test both on your own handset.

What actual phones record

We do not publish estimated speeds. TokForge has a benchmark harness built into the app, and results submitted from it land on a public board carrying the device, the SoC, the model, the backend, the quantization, the app version, the peak memory and the date. Devices on the board range from budget handsets to current flagships, and results are grouped by model size class so you can compare like with like.

If you want to know what a specific model does on hardware close to yours, that board is the honest answer, and it stays current in a way a table on this page could not.

Open the phone-speed leaderboard

Bringing your own model

You are not restricted to the curated list. TokForge browses and downloads from Hugging Face inside the app, and imports GGUF and MNN files you already have. Each download shows its license and model card first, so you know what you are agreeing to before the file lands on your phone.

Common questions

How do I know if a model will fit before downloading it?

TokForge compares each model against your device's actual free memory and marks the ones that will not fit. It will still let you try if you insist, but it tells you first.

Can I keep several models installed?

Yes. Storage is the only limit, and models can be deleted and re-downloaded freely at any time.

Which model should I start with?

Take the largest size your RAM tier comfortably supports from the table above, in a Q4-class quantization, and use it for a day before changing anything. Most people over-tune before they know what they actually need.

Do bigger models always answer better?

Usually, but not always, and on a phone the trade is real: a bigger model is slower, hotter, and more likely to be evicted from memory when you switch apps. A 4B model you actually use beats a 14B model you avoid because it is sluggish.

Keep reading

• Best AI models for your phone: a size guide
• Is your phone fast enough for local AI?
• How to run AI offline on your Android phone
• Vulkan vs OpenCL vs CPU: which backend and when