FreeLLMFree AI models on your phone
A phone can run a small AI model offline, with no account and no internet. What fits depends on its memory (RAM). Here is what to expect, and which free apps to use.
What fits in your phone’s memory
Rule of thumb: the model can use about 40% of the phone’s RAM; the system and your apps need the rest. Sizes are 4-bit files from the labs’ official Hugging Face pages, ordered by downloads (popularity, not quality).
8 GB of RAM · Most phones from 2023 on · models up to about 3.2 GB
| Model | Size | 4-bit file | Downloads/month |
|---|---|---|---|
| Qwen3-0.6B Qwen | 0.8 B | 0.9 GB | 29.5M |
| Qwen3-4B Qwen | 4 B | 2.8 GB | 11.8M |
| gemma-3-1b-it | 1 B | 1.0 GB | 1.6M |
| NVIDIA-Nemotron-3-Nano-4B-BF16 nvidia | 4 B | 2.8 GB | 1.6M |
| SmolVLM2-500M-Video-Instruct HuggingFaceTB | 0.5 B | 0.7 GB | 1.4M |
| MiniCPM5-2B openbmb | 2.5 B | 1.9 GB | 1.3M |
| DeepSeek-R1-Distill-Qwen-1.5B deepseek-ai | 1.8 B | 1.5 GB | 1.3M |
| SmolLM3-3B HuggingFaceTB | 3.1 B | 2.3 GB | 749K |
12 GB of RAM · Recent flagships · everything above, plus models up to about 4.8 GB
| Model | Size | 4-bit file | Downloads/month |
|---|---|---|---|
| Qwen3.5-4B Qwen | 4.7 B | 3.2 GB | 7.8M |
| gemma-3n-E2B-it | 5.4 B | 3.7 GB | 119K |
| OLMoE-1B-7B-0125-Instruct allenai | 6.9 B | 4.6 GB | 106K |
| Molmo2-ER allenai | 4.9 B | 3.4 GB | 28K |
| Apriel-5B-Instruct ServiceNow-AI | 4.8 B | 3.3 GB | 3K |
16 GB of RAM · Top phones and tablets · everything above, plus models up to about 6.4 GB
| Model | Size | 4-bit file | Downloads/month |
|---|---|---|---|
| Qwen3-8B Qwen | 8.2 B | 5.4 GB | 8.9M |
| Qwen3.5-9B Qwen | 9.7 B | 6.3 GB | 7.5M |
| Apertus-v1.5-8B swiss-ai | 8.9 B | 5.8 GB | 524K |
| Olmo-3-7B-Think allenai | 7.3 B | 4.8 GB | 489K |
| DeepSeek-R1-0528-Qwen3-8B deepseek-ai | 8.2 B | 5.4 GB | 445K |
| Apertus-8B-Instruct-2509 swiss-ai | 8.1 B | 5.3 GB | 434K |
| Olmo-3-7B-Instruct allenai | 7.3 B | 4.8 GB | 400K |
| DeepSeek-R1-Distill-Qwen-7B deepseek-ai | 7.6 B | 5.0 GB | 386K |
The exception: MiniCPM-V 4.7 with expert streaming
MiniCPM-V 4.7 35B-A3B by OpenBMB has 35.2 billion parameters but uses only about 3 billion for each word (a mixture of experts). Its makers keep the experts in the phone’s storage and load them just before they are needed, so only about 3 GB sit in memory. They report more than 13 tokens per second on an iPhone 17 Pro Max (model page). It reads images, video and documents too.
What this does not mean: the whole model still needs about 22 GB of storage, it needs their special build of llama.cpp, and speed depends on how fast the phone’s storage is. On a graphics card the same model is much faster: see its page.
Free apps to run them
- PocketPal AI: open-source app that downloads GGUF models from Hugging Face and runs them offline (MIT license).
- Google AI Edge Gallery: Google’s open-source app to try on-device models, including Gemma (Apache-2.0).
- MLC LLM: engine with a chat app for phones; more technical (Apache-2.0).
Download the model inside the app (look for a 4-bit or Q4 version), then it works offline. Bigger models drain the battery faster and make the phone warm.
Need more?
For bigger and smarter models, use them free online (leaderboard), try them without an account in Free Arena, or see what a graphics card can run in GPUs for local AI.