Qwen3.5 122B-A10B

Qwen3.5-122B-A10B is published by the Qwen organization on Hugging Face as an open-weights base model, with its name and repository path appearing in third-party references to the family. The 122B-A10B designation suggests a very large parameter footprint paired with a smaller active parameter count per token, a naming convention associated with mixture-of-experts architectures, but the supplied excerpts do not provide direct documentation of the modeling approach or training pipeline. Its positioning as a base release makes it suitable for downstream adaptation, including research forks and community fine-tuning efforts rather than turnkey deployment.

The strongest practical signal comes from the existence of a community derivative, wangzhang/Qwen3.5-122B-A10B-abliterix, which treats Qwen/Qwen3.5-122B-A10B as its upstream parent. That derivative describes itself as the largest abliterated model in the Qwen3.5 line, indicating that the base is large enough to be a meaningful target for capability-preserving fine-tuning work. For practitioners, this points to a base model that invites experimentation through derivative releases while the upstream itself remains the reference point researchers and tinkerers build upon.