Qwen3.5 122B-A10B

Qwen3.5 122B-A10B is part of the Qwen family of foundation models, with the base weight release published on the Qwen organization's Hugging Face page under the path Qwen/Qwen3.5-122B-A10B. A community derivative hosted by user wangzhang explicitly identifies the source repository, which confirms the upstream model's public availability. The "A10B" designation and "122B" total parameter footprint indicate a Mixture-of-Experts design with a large total parameter count and a smaller active subset per token, a pattern consistent with MoE architectures that balance capacity and inference cost within the model.

Independent reporting through DeepInfra's April 2026 benchmark blog post treats Qwen3.5 122B-A10B as a multimodal vision-language foundation model supporting text, image, and video inputs and producing text outputs. The base model is positioned as a mid-tier multimodal entry within the Qwen3.5 lineup, designed for general-purpose vision-language tasks that combine language understanding with visual and video comprehension within the model. The combination of MoE efficiency and multimodal input support makes it suited for applications such as document and chart reasoning, video question answering, and tool-augmented workflows that need both visual grounding and language generation.