A team named Neykosa has benchmarked a Qwen3 0.6B model running on a budget Android phone, the Galaxy A05s with 4 GB of RAM. The model achieved a speed of 14.8 tokens/second with an empty context and 11.0 tokens/second with a 768-token context. These measurements were taken using llama.cpp and reported through Firebase Test Lab, highlighting that context depth impacts throughput even with the same model and runtime. The team also conducted separate whole-app tests to measure answer times, emphasizing the importance of distinguishing between raw speed benchmarks and actual application response times. AI
IMPACT Demonstrates the feasibility of running LLMs on low-resource mobile devices, potentially enabling new on-device AI applications.
RANK_REASON Research benchmark of a small LLM on consumer hardware. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →