Qwen 3.5 0.8B

Qwen 3.5 0.8B is the smallest-scale model in the Qwen 3.5 family from Alibaba's Qwen team. At roughly 0.8 billion parameters, it is intended for low-resource settings such as edge devices, local inference and fine-tuning or research experiments. The "Qwen 3.5B" alias sometimes used for it is likely inaccurate; the model discussed is the 0.8B variant.

The available material says little about its specifications. A YouTube video by Matthew Berman covered a research paper (referred to as "Astra") that used this model in its final experiment. The method edited the model's weights to reduce looping behaviour. The video reports MATH 500 rising from under 8% to over 48% and GSM8K from 13.8% to 67.1%. These figures come from a creator's summary and have not been checked against the paper. They also show that the base model loops heavily on reasoning tasks.

Qwen models are generally distributed through the Qwen website and chat app, Alibaba Cloud APIs, and Hugging Face/GitHub for open-weight releases. Exact license, context length and release date for this specific checkpoint were not in the provided pages, so they are left unspecified here.