Jackrong/Qwopus3.6-35B-A3B-Coder
Jackrong/Qwopus3.6-35B-A3B-Coder is a 35.1 billion parameter hybrid sparse Mixture-of-Experts (MoE) model, based on the Qwen3.6-35B-A3B architecture, with approximately 3 billion active parameters per token. Developed by Jackrong, this model is specifically fine-tuned for agentic coding workflows, emphasizing execution efficiency, lower token waste, and stable behavior without relying on explicit long reasoning. It excels at multi-turn code tasks, tool calling, and debugging, making it ideal for local high-frequency coding applications and agent harnesses.
Loading preview...
Qwopus-3.6-35B-A3B-Coder: An Agentic Coder Model
Qwopus-3.6-35B-A3B-Coder is a specialized 35.1 billion parameter hybrid sparse Mixture-of-Experts (MoE) model, built upon the Qwen3.6-35B-A3B foundation. It features approximately 3 billion active parameters per token, making it efficient for local coding workflows. The model's core objective is to optimize for execution efficiency in agentic coding, rather than verbose, visible reasoning.
Key Capabilities & Differentiators
- Thinking-Off Execution: Designed to perform agent work with fewer tokens, faster turns, and more stable tool behavior by minimizing explicit long thinking steps.
- Token Efficiency: Reduces unnecessary long-form reasoning in routine implementation steps, leading to lower token waste and latency.
- Workflow Stability: Maintains state and task focus across multi-turn code tasks, file edits, tool calls, and retries.
- Agent Harness Compatibility: Aims to fit various agent harness styles, including Codex-style, OpenHands-style, Claude Code-style, and OpenCode-style.
- SWE-bench Performance: Achieved a 62.4% score on a 300-case SWE-bench run using a Q5_K_M quantized build with thinking disabled, demonstrating practical execution capabilities.
- Behavioral Strengths: Outperforms comparative models in categories like legit-request compliance, integrity under pressure, multi-turn orchestration, large code deliverables, and sustained debugging in "thinking-off" mode.
Recommended Use Cases
- Agentic Coding Workflows: Ideal for Codex-style, OpenHands/OpenCode coding loops, and local tool-calling agents.
- Debugging & Patching: Excels at repository-level debugging, multi-file patch generation, and automated test-fix cycles.
- Large Context Codebase Tasks: Suitable for tasks requiring practical execution quality over verbose reasoning, such as DevOps scripting and code review assistance.
This model is an experimental community release, primarily intended for research and local coding-agent evaluation, with a focus on practical, high-frequency coding tasks.