Positioned as an experimental preview of the Qwen4 architecture lineage, this 125B-parameter sparse model activates roughly 6B parameters per token through a hybrid-attention mixture-of-experts design, paired with a vision encoder for image and video understanding. The architecture targets coding and agent-style workloads, with open weights making it suitable for self-hosting and downstream fine-tuning rather than purely API-mediated experimentation.

In coding evaluations the model posts competitive but not state-of-the-art results: SWE-Bench Pro at 62.5%, SWE-Bench Multilingual at 81%, DeepSWE 1.1 at 58.7%, and NL2Repo-Bench at 48.1%, placing it within the upper tier of tested systems without leading any single benchmark. Its very large context budget supports repository-level reasoning and long multimodal sessions, while per-million-token pricing remains low, making it a practical choice for developers who want an open-weight, multimodal helper for long-context code and agent pipelines.