Qwen3.5-0.8B

Qwen3.5-0.8B is a compact 0.8B-parameter causal language-vision model from the Qwen family (Alibaba). It has 24 layers and a hybrid architecture combining Gated DeltaNet with a sparse Mixture-of-Experts (MoE), optimized for efficient inference.

The model is multimodal — accepting text, images, and video — and supports a native context of up to 262,144 tokens. It runs in non-thinking mode by default, but a thinking (reasoning) mode can be enabled via API parameters; it also supports tool calling and agent applications.

Qwen3.5-0.8B underwent both pre-training and post-training, targeting high utility across reasoning, coding, and visual understanding. As the smallest variant of the series it performs more modestly than the larger 2B/4B models, but stays very lightweight. Released under the Apache 2.0 license (for research and commercial use).