Jackrong/Qwopus3.5-27B-v3.5

VISIONConcurrency Cost:2Model Size:27BQuant:FP8Ctx Length:32kTool Calling:SupportedPublished:Apr 12, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Cold

Qwopus3.5-27B-v3.5 is a 27 billion parameter language model developed by Jackrong, based on Qwen3.5-27B. This version is a data-scaled continuation of Qwopus3.5-27B-v3, trained with approximately twice the SFT data to enhance generalization across domains like mathematics, programming, and multi-turn interactions. It is specifically designed for structured reasoning, tool-augmented workflows, and multi-step agentic tasks, showing improved performance in complex problem-solving and coding scenarios.

Loading preview...

Overview

Jackrong's Qwopus3.5-27B-v3.5 is a 27 billion parameter model built upon Qwen3.5-27B, representing a significant data-scaled evolution from its predecessor, Qwopus3.5-27B-v3. This version focuses on enhancing generalization by expanding its training data to include a broader array of domains such as mathematics, programming, puzzle-solving, multilingual dialogue, and STEM-related tasks. Unlike previous iterations, v3.5 does not introduce new architectural changes, RL stages, or template redesigns, instead relying on a substantial increase (approximately 2x) in high-quality Supervised Fine-Tuning (SFT) data.

Key Capabilities

  • Enhanced Structured Reasoning: Designed for complex problem-solving, leveraging structured reasoning over simple Chain-of-Thought (CoT) mimicry.
  • Tool-Augmented Workflows: Optimized for integration with external tools, supporting more effective and reliable use of existing knowledge.
  • Multi-Step Agentic Tasks: Excels in scenarios requiring multiple steps and agentic planning, particularly in coding and diagnostic tasks.
  • Improved Coding Performance: Demonstrated significant gains in SWE-style capability tests, including code inspection, bug diagnosis, and action planning, successfully passing 14 out of 15 programming tasks in a 44-case suite.
  • Data-Driven Generalization: Improvements are attributed to scaling high-quality SFT data, leading to better utilization and activation of latent knowledge.

Good For

  • Developers requiring robust structured reasoning for complex logical problems.
  • Applications involving tool use and agentic workflows.
  • Coding tasks, especially those requiring multi-step problem-solving, bug diagnosis, and code generation.
  • Use cases demanding strong performance in mathematics, programming, and STEM-related domains.

Limitations

  • Potential for overfitting if data scaling exceeds optimal regimes.
  • Reasoning may still exhibit instability in certain edge cases.
  • Tool-calling performance is dependent on the specific environment integration.
  • Not all capabilities have been fully benchmarked yet.

Finetunes
2 models