MonoWare
MonoWare
Development journal
A lighter place for product thoughts, release notes, and small lessons from the workbench.
General notes
A controlled comparison of 10 compact GGUF models, including Llama 3.2, SmolLM, and LFM2.5, with iPhone 4K/8K resource data and a 96-case bilingual Agent test.
A focused iPhone 15 Pro Max test of Qwen3.5-2B with 3,584- and 7,680-token prompts, measuring prefill, TTFT, generation, memory, and thermal state.
Our phone-local model test design, with all 96 bilingual cases and expected routes, structured metrics, fixed decoding, iPhone resource runs, evidence, and limits.