Qwen and Fable: An open-weights agentic coding model. 35B Mixture-of-Experts (huggingface.co)

The AI/ML community has welcomed the announcement of Qwable-v1, an open-weights agentic coding model with a staggering 35 billion parameters and a sophisticated 3 billion active mixture-of-experts architecture. Built by layering behavior from Claude Fable-5 onto a distilled reasoning model from Qwen3.6-35B, Qwable-v1 allows enhanced tool-use capabilities while maintaining strong reasoning skills. The model operates using XML-formatted tool calls for tasks such as file edits and running commands, accessible through agent-style prompts, making it particularly valuable for developers and researchers needing efficient coding assistance. The significance of Qwable-v1 lies in its unique training approach and technical flexibility. The model employs a chained distillation process, incorporating reasoning and tool-use traces effectively, while also being designed to run on standard consumer GPUs with sufficient memory. However, its sustainability is uncertain due to the suspension of Claude Fable-5 under U.S. export controls, which may limit future enhancements. As the community monitors potential uploads or modifications to the training corpus, the ongoing updates are expected to enhance the model's performance, particularly in agentic coding scenarios.