ProCreations/grug-27b
ProCreations/grug-27b is a 27 billion parameter language model based on the Qwen3.6-27B architecture, developed by ProCreations. It is specifically fine-tuned to produce highly condensed, efficient internal reasoning ("grug think") while maintaining high-quality, standard English outputs. This model excels at complex problem-solving tasks, particularly in code generation and mathematical derivations, by significantly reducing token expenditure for internal thought processes. It offers substantial token savings and improved stability compared to its base model, making it ideal for agentic workflows and hard reasoning tasks.
Loading preview...
Overview
ProCreations/grug-27b is a 27 billion parameter model built upon the Qwen3.6-27B base, developed by ProCreations. Its core innovation lies in its "grug think" mechanism, which enables highly efficient internal reasoning using significantly fewer tokens than traditional models, while still generating high-quality, standard English responses. This model was trained using LoRA and a unique "think-only loss" approach on agent trajectories and fresh, high-quality data, including hard math derivations and code design.
Key Capabilities
- Efficient Reasoning: Achieves up to 98% token reduction for internal thought processes compared to its base model, without compromising answer quality.
- Robustness: Engineered to prevent repetitive loops and "think-sickness" observed in other models, even under stress tests.
- Adaptive Thinking: Dynamically adjusts the depth of its internal reasoning based on task complexity, providing concise thoughts for easy tasks and detailed deliberation for hard ones.
- Strong Performance: Maintains or improves upon base model benchmarks in coding (HumanEval, MBPP) and math (GSM8K, MATH-500), with notable gains in tool use accuracy.
- Prompt Stability: Demonstrates consistent performance across varying system prompts, unlike its base model which can be highly sensitive.
Good For
- Agentic Workflows: Optimized for use in agent frameworks where efficient internal reasoning and tool calling are critical.
- Code Generation & Problem Solving: Excels at complex coding tasks and mathematical derivations, providing high-quality solutions with reduced computational overhead.
- Resource-Constrained Environments: Its token efficiency makes it suitable for applications where minimizing token usage is important.
- Developers Seeking Stability: Offers a more stable and predictable output generation experience compared to its base model.