Logics-MLLM/Logics-SWE-Qwen3.6-27B

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 18, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Logics-SWE-Qwen3.6-27B is a 27 billion parameter model developed by Logics-MLLM, based on Qwen3.6-27B, specifically designed for repository-level software engineering agents. It integrates category-aware expert development with multi-teacher on-policy distillation to enhance performance on complex coding tasks. This model excels at navigating codebases, editing files, executing commands, and iteratively repairing solutions, making it highly effective for automated software development workflows. It demonstrates improved task resolution on benchmarks like Pro-618 and SWE-bench Multilingual compared to its base model.

Loading preview...

Logics-SWE-Qwen3.6-27B Overview

Logics-SWE-Qwen3.6-27B is a 27 billion parameter model developed by Logics-MLLM, specifically engineered for repository-level software engineering agents. It builds upon the Qwen3.6-27B base model, incorporating a novel approach that combines category-aware expert development with multi-teacher on-policy distillation (MOPD) to create a single, highly capable deployment policy.

Key Capabilities & Innovations

  • Repository-Level Task Resolution: Designed to handle complex software engineering tasks requiring navigation, editing, execution, and iterative repair within code repositories.
  • Refresh–Repair–Expand (RRE) Expert Development: Utilizes a three-stage process (Refresh, Repair, Expand) for expert development, alternating executable-reward Reinforcement Learning (RL) with supervised replay of successful trajectories across three distinct software engineering categories (service/data/security, user-facing apps, systems/tooling).
  • Multi-Teacher On-Policy Distillation (MOPD): Integrates learned behaviors from multiple experts into a single student model, where the student generates its own trajectories and receives token-level supervision from category experts, without direct environment rewards during distillation.
  • Enhanced Performance: Achieves significant improvements in mean task resolution, with a +5.40 percentage point increase on Pro-618 and +2.78 percentage point increase on SWE-bench Multilingual compared to the base model.

Intended Use Cases

This model is ideal for research and development in:

  • Repository-level issue resolution
  • Advanced coding agents
  • Long-horizon post-training strategies
  • Expert system development
  • Policy distillation techniques

Finetunes
1 models