As a smaller sibling within the Qwen3.6 family, this release is built as a sparse Mixture-of-Experts model with 35 billion total parameters but only around 3 billion active parameters per token, an efficiency-oriented design that lowers compute and memory demands without sacrificing capability. The open-weights distribution lets developers run the model locally or self-host it, with quantization-friendly GGUF builds that fit on consumer hardware such as 24GB-RAM Macs. It carries forward the lineage of Qwen3.5 while targeting coding and tool-driven workflows, positioning it as a practical option for repository-level reasoning and multi-step agentic tasks rather than as a general-purpose chatbot.
The model is tuned for agentic coding, combining a very large 262,144-token context window with native tool calling and reasoning parsing, which makes it well suited for long codebases, multi-file edits, and orchestrated developer assistants. Independent benchmark coverage highlights strong coding results, including a 73.4 score on SWE-bench Verified and a 51.5 score on Terminal-Bench 2.0, and notes that it can outperform dense models in its weight class while remaining competitive with much larger frontier systems. Recommended vLLM deployments use single-node tensor parallelism with features like auto tool-choice and reasoning-mode decoding, supporting flexible serving across a wide range of data-center GPUs as well as high-end workstation cards.
