reasonwang/SkillGym-Qwen3.5-27B

VISIONPricing:Input $1.6 / Cached $0.15 / Output $12Concurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Sep 22, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

SkillGym-Qwen3.5-27B is a 27 billion parameter language model developed by Reason Wang, fine-tuned from Qwen3.5-27B. It is specifically trained on SkillGym trajectories to enable agents to effectively use predefined skills, which are folders containing instructions, reference documents, and scripts. This model excels at applying relevant skills through tool calls or shell commands within a sandboxed environment, making it ideal for agentic task execution rather than general chat assistance. It features a 32768 token context length and is optimized for skill-use scenarios.

Loading preview...

Overview

SkillGym-Qwen3.5-27B is a 27 billion parameter model, fine-tuned from Qwen3.5-27B by Reason Wang. Its primary purpose is to enable AI agents to effectively utilize "skills" – structured sets of instructions, documents, and scripts – to solve tasks. The model is trained to interpret these skills and apply them through tool calls or shell commands within a sandboxed environment, as detailed in the associated paper, "SkillGym: Training Skill-Use Agents with Automatic Verifiable Environment Generation."

Key Capabilities

  • Agent Skill Utilization: Proficiently reads and applies agent skills, which are collections of instructions and scripts.
  • Tool and Shell Command Execution: Capable of making tool calls and executing shell commands based on skill instructions.
  • Task Success: Demonstrates improved task success rates on skill-based benchmarks compared to its base model.
  • Reasoning Traces: Trained with reasoning traces to enhance decision-making processes.

Performance Highlights

This model shows significant improvements in skill-use benchmarks:

  • SkillGym: Achieves 62.8, outperforming the base Qwen3.5-27B (55.3).
  • SkillEval: Scores 80.1 ± 1.5, compared to 76.2 ± 1.1 for the base model.
  • Skill-Use-Bench: Reaches 74.2, a substantial increase from the base model's 28.3.

Good for

  • Developing AI agents that require the ability to interpret and execute predefined skills.
  • Automating complex tasks that involve tool usage and shell command execution.
  • Research and development in agentic AI systems and skill-based task completion.

Limitations

  • Not a General Chat Assistant: This model is specifically tuned for agent policies and skill-use, not for general conversational AI.
  • Context Budget: Long task episodes may exhaust the 32768 token context length.
  • Reasoning Repetition: Occasional repetition in reasoning and hitting per-turn token limits can occur.

Finetunes
1 models