ProCreations/grug-v1.1-qwen-3.8-27b

VISIONConcurrent Unit Cost:2Model Size:27BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Aug 17, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

ProCreations/grug-v1.1-qwen-3.8-27b is a 27 billion parameter language model based on Qwen3.8, developed by ProCreations. This model is specifically fine-tuned to significantly reduce reasoning token usage while maintaining high answer quality, making it highly efficient for agentic tasks and tool use. It excels at picking the correct tool and managing agentic workflows with substantially fewer internal 'thought' tokens compared to its base model.

Loading preview...

Overview

ProCreations/grug-v1.1-qwen-3.8-27b is an instruction-tuned model built upon the Qwen3.8-27B base, designed to optimize token efficiency in reasoning while preserving high-quality output. It achieves this by stripping away grammatical padding from its internal thought processes, resulting in significantly fewer 'thinking' tokens per answer. This model is an iteration over the previous grug v1, with key improvements in tool selection and repetition handling.

Key Capabilities & Differentiators

  • Extreme Token Efficiency: Reduces reasoning token usage by 7-30x compared to the base Qwen3.8 model, particularly notable in agentic steps (20 tokens vs. 108.5) and HumanEval (79.5 tokens vs. 559).
  • Superior Tool Use: Achieves 97.1% accuracy in picking the right tool, a substantial improvement over the base model's 23.5%.
  • Robustness: Demonstrates strong performance in repetition stress tests (88.4%) and ensures 100% closed 'think' blocks, preventing loops.
  • Configurable Reasoning Effort: Supports reasoning_effort settings (low, medium, xhigh), with medium being the recommended and tuned setting for optimal balance between performance and token cost.
  • Targeted Fine-tuning: Developed through SFT on a 1M-row 'grug corpus' and corrective LoRA applied at a precise 0.5x strength to prevent overfitting and maintain code generation quality alongside tool use.

Good For

  • Agentic Workflows: Ideal for applications requiring precise tool calling and efficient multi-step reasoning.
  • Cost-Sensitive Applications: Significantly reduces token expenditure for internal reasoning, leading to lower operational costs.
  • Code Generation & Problem Solving: Maintains strong performance on benchmarks like HumanEval (94.5%) and MBPP (88.0%) while being more efficient.
  • Developers seeking optimized LLM behavior: Especially those who need a model that 'thinks' efficiently without verbose internal monologues.

Finetunes
8 models