kai-os/Grug-12B

Grug 12B: Compact Reasoning Model

Grug 12B is a 12 billion parameter model fine-tuned from google/gemma-4-12B-it with a focus on compact reasoning. Its primary goal is to generate shorter, denser, and less verbose reasoning traces compared to its base model, while maintaining or improving answer quality.

Key Capabilities & Features

• Efficient Reasoning: Trained to produce high-density reasoning steps, minimizing filler words and explicitly preserving key constraints, branching decisions, invariants, edge cases, and final-answer checks.
• Token-Efficiency: Aims for lower reasoning-token usage, making it potentially more cost-effective for applications where token count is a concern.
• Targeted Fine-tuning: Utilizes a specialized dataset derived from verbose reasoning traces, transformed into compact versions using cyankiwi/Qwen3.6-35B-A3B-AWQ-4bit.
• Performance: A small local evaluation showed Grug 12B achieving 100% proxy accuracy on math problems with significantly fewer generated tokens (68.94 avg) compared to the base model (228.53 avg).

Training Details

The model was trained using QLoRA/PEFT LoRA, with 4-bit NF4 quantization and BF16 compute, on a dataset of 5,166 compact reasoning SFT rows. The training data was carefully filtered for licenses and excluded sources from specific commercial LLMs.

Good For

• Applications requiring token-efficient reasoning.
• Tasks like mathematical problem-solving and code debugging where concise, high-density reasoning is beneficial.
• Use cases where reducing verbosity in reasoning steps is a priority without compromising accuracy.