DownloadQwable-v1 is an open-weight agentic coding model created through a chained distillation process based on Qwen3.6-35B-A3B. The model combines two distinct training stages: first, it was fine-tuned on reasoning traces derived from Claude Opus 4.7 to improve structured reasoning, and then further trained on Claude Fable-5 agentic tool-use traces to develop autonomous coding and tool-calling behavior. The result is a 35B Mixture-of-Experts model with only 3B active parameters that can switch between deep reasoning and agent-style execution depending on the system prompt. In normal conversations, it behaves like a reasoning-focused assistant that generates explicit <think> chains before answering. When configured as an agent, it can emit structured tool-use XML for file editing, shell commands, codebase navigation, and workflow automation. Qwable-v1 is designed specifically for software engineering, code editing, debugging, and autonomous coding workflows.
Features
- 35B Mixture-of-Experts architecture with 3B active parameters
- Built on Qwen3.6-35B-A3B as the foundation model
- Distilled from Claude Opus 4.7 reasoning traces
- Enhanced with Claude Fable-5 agentic tool-use behavior
- Explicit think reasoning chains before responses
- Structured XML tool calling for coding agents
- Optimized for software engineering and codebase workflows
- Runs locally through GGUF quantizations and consumer GPUs
Project Samples
