0ronoCris/Affine-5dfqbbh8ev-qwopus
0ronoCris/Affine-5dfqbbh8ev-qwopus is a 35.1 billion parameter coding-agent model, built on the Qwopus3.6-35B-A3B (Qwen3.6-35B-A3B) hybrid sparse MoE architecture with approximately 3 billion active parameters per token. This model is specifically fine-tuned for execution efficiency in agentic coding workflows, prioritizing faster decisions, lower token waste, and stable multi-turn behavior over explicit long reasoning chains. It excels at tasks like code editing, debugging, and tool-use orchestration within agent harnesses, making it suitable for local high-frequency coding tasks.
Loading preview...
What is 0ronoCris/Affine-5dfqbbh8ev-qwopus?
This model, named Qwopus-3.6-35B-A3B-Coder, is a 35.1 billion total parameter (3B active parameters) coding-agent fine-tune based on the Qwen3.6-35B-A3B MoE architecture. Its core objective is to optimize for execution efficiency in agentic coding workflows, rather than generating verbose reasoning. It is designed for scenarios where a coding agent repeatedly performs operational steps like reading files, choosing tools, editing code, running tests, and reacting to errors, aiming for less token waste, lower latency, and more stable behavior when explicit long thinking is disabled.
Key Capabilities & Optimizations
- Faster Agent Loops: Optimized for rapid tool decisions, code patching, test runs, and error-driven debugging without requiring extensive deliberation at each step.
- Token Efficiency: Reduces unnecessary long-form reasoning, leading to lower token consumption in routine implementation tasks.
- Workflow Stability: Maintains state and keeps multi-turn code tasks on track across various operations like file edits and tool calls.
- Local Deployment Fit: Designed to make high-frequency coding tasks practical on local or self-hosted inference setups.
- SWE-bench Performance: Achieved 62.4% on a 300-case SWE-bench run (submitted patches, thinking-off mode, Q5_K_M quantized model), demonstrating strong practical execution.
When to Use This Model
This model is a strong fit for:
- Agentic Coding Workflows: Specifically designed for Codex-style, OpenHands, Claude Code, and OpenCode-style agent harnesses.
- Repository-level Tasks: Ideal for long-running repository edits, automated debugging, and multi-file patch generation.
- Automated Test-Fix Cycles: Excels in iterative processes of testing, identifying errors, and applying fixes.
- Local Tool-Calling Agents: Optimized for scenarios requiring low-latency, high-frequency tool interactions on local hardware.
It is particularly useful when practical execution quality and token efficiency are prioritized over verbose, explicit reasoning.