Team Neykosa has benchmarked a 0.6B parameter model, Qwen3 0.6B, running on a 4GB Galaxy A05s phone. The model achieved a speed of 14.8 tokens per second using llama.cpp, with specific settings including 6-bit Q6_K quantization and 4 threads. However, the team emphasizes that this speed metric does not reflect the user experience, which is more dependent on overall response time. Separate tests for their Nesayo offline edition showed response times ranging from 4.25 seconds for a stock answer to 18.8 seconds for the first launch, with model memory usage around 673 MB. AI
IMPACT Demonstrates the feasibility of running smaller LLMs on low-resource mobile devices, potentially enabling offline AI applications.
RANK_REASON Benchmark results for a small LLM running on a mobile device. [lever_c_demoted from research: ic=1 ai=1.0]
AI-generated summary · Google Gemini · from 1 sources. How we write summaries →