Qwen3-0.6B is the roughly 0.6B-parameter small model in Alibaba's Qwen3 family, often used for local and low-resource experiments. Recent community work converted it into a Mixture-of-Experts model with GGUF releases, trading some quality for faster generation; the University of Waterloo's ProgramAsWeights uses it as a frozen interpreter for generated LoRA programs; and a developer ran it in Q4_K_M via llama.cpp on a 2017 Galaxy Note 8 to drive desktop Chrome.
Community converts Qwen3-0.6B into MoE modelA developer converted Qwen3-0.6B into a Mixture-of-Experts model and released GGUF versions. Lacking Qwen3's shared expert, quality dropped; a pseudo-shared expert recovered some of it. MoE perplexity and task scores were worse than dense, but generation speed rose sharply.
ProgramAsWeights Compiles English into Local Neural ProgramsResearchers at the University of Waterloo released ProgramAsWeights, an open-source project that compiles English function descriptions into reusable neural programs that run locally. A finetuned Qwen3-4B generates LoRA adapters for a frozen Qwen3-0.6B interpreter, separating compilation from inference so programs can be saved, distributed and composed with ordinary code.
Qwen3-0.6B drives real desktop Chrome on old phoneA developer ran Qwen3-0.6B (Q4_K_M) with llama.cpp on a Galaxy Note 8 (2017, 6GB) and used a relay to drive a real desktop Chrome. The model performed structured page-perception tasks and scored 10/10 on several tests. The experiment shows structured browser perception was far more efficient than feeding raw HTML.
The A.I. Daily