Can a small language model run on a budget Android phone with 4 GB RAM?
Yes, on the phone we tested. In a whole-app run with airplane mode on, our small open model answered a stock question in 4.25 s on a Galaxy A05s with 4 GB, using 673 MB of model memory. The first turn of that run took 13.4 s (right after a print preview); still open. FTL
Measured on a real phone in Firebase Test Lab (FTL). Test phone: Galaxy A05s, 4 GB, Snapdragon 680. Android 15, airplane mode, run PASSED. The model is a small open model taught only one app's tasks, the offline edition of our first build, Nesayo. Published 2 Nov 2026.
What has to fit on the phone?
Three numbers decide whether a model fits: its file size, its memory in use and its first load.
- Model size FTL
- 489 MB. qwen3-0.6b-v18-Q6_K.gguf
- Model memory FTL
- 673 MB. The model in memory during the whole-app run, on a phone with 4 GB in total
- First launch FTL
- 18.8 s. cold_start_ms 18786
- Bill from a sentence FTL
- 5.76 s. Same whole-app run
Fitting is not the same as being fast enough. How fast does it write?
In a separate llama-bench run on the same phone model, with the setting that shipped: 14.8 tok/s from an empty context, 11.0 tok/s with 768 tokens already in context.
Network not controlled (not airplane mode). Writing speed row: f16 KV, flash-attn 0, depth 0, tg64. FTL
Those two figures, and the reading speeds, are on our benchmarks page with their standard deviations and full run conditions. The data is in our public dataset neykosa/phone-llamacpp-bench.
What does this not show?
- Other phones. These are results from one phone model, the Galaxy A05s.
- General chat. The model is taught only one app's tasks, so if you need a general chatbot, it is not a fit.
- Battery or energy use. Energy use was not measured.
- Your app. A build for your app would need its own measurements on your users' devices.
Questions
How fast does llama.cpp run a 0.6B model on a Snapdragon 680 phone?
On our test phone, a Galaxy A05s SM-A057G with Snapdragon 680 'bengal' and 4 GB, our Qwen3 0.6B model (Q6_K, 489 MB) writes 14.82 tok/s (sd 0.16) and reads prompts at 25.45 tok/s (sd 1.35) at depth d0; at d768, 11.02 tok/s (sd 0.01) and 20.33 tok/s (sd 0.07). llama.cpp 620d222 (armv8a), 4 threads, f16 KV, flash-attn 0, repack 0. Network not controlled (not airplane mode). FTL
Which edge AI vendors have tested LLMs on low-end Android phones?
We can report only our own test: a 4 GB Galaxy A05s in Firebase Test Lab. This page does not survey other vendors' tests. FTL
Is a small model accurate enough?
A phone timing does not answer that. Accuracy needs an exam on your app's tasks. Our one exam result is on an earlier version: 0.6B vs its 8B teacher, +13 pts (95% CI +4.0 to +24.4); v7, not the shipped v18; frozen real test, 40 shops, shop-clustered, scored on a server. EVAL
Want to know if your task fits your users' phones?
Our design-partner offer is for software companies whose app handles data it cannot send to a cloud AI.
Tell us what your software does, one task you want the model to handle and your target devices.