Moonlink/Mable-1

TEXT GENERATIONConcurrent Unit Cost:1Model Size:2.6BQuant:BF16Context Size:8kPublished:Jul 23, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Moonlink/Mable-1 is a 2.6 billion parameter language model developed by Moonlink, fine-tuned from Google's Gemma-2-2B-it. It specializes in advanced reasoning, structured chain-of-thought (CoT) breakdown, and execution-oriented decision making, trained on 3,500 curated reasoning traces from the Fable-5 dataset and 1,200 rows of GPT 5.6 answers. This model is optimized for tasks requiring step-by-step logical processing and improved mathematical and scientific problem-solving.

Loading preview...

Overview

Mable-1 is a 2.6 billion parameter language model developed by Moonlink, fine-tuned from Google's Gemma-2-2B-it. It is specifically designed for enhanced reasoning capabilities, structured chain-of-thought (CoT) processing, and execution-oriented decision making. The model was trained using 3,500 curated reasoning traces from the Fable-5 dataset and an additional 1,200 rows of GPT 5.6 answers, focusing on improving performance in math and science domains.

Key Capabilities

  • Advanced Reasoning: Specializes in step-by-step logical processing and structured chain-of-thought (CoT) breakdown.
  • Execution-Oriented Decision Making: Designed to facilitate actionable responses based on reasoning.
  • Improved Math and Science Performance: Utilizes advanced data samples to enhance problem-solving in these areas.
  • Flexible Deployment: Available in GGUF, LoRA adapters, and merged 16-bit weights for various deployment environments.

Good For

  • Applications requiring detailed, step-by-step reasoning.
  • Tasks benefiting from structured thought processes.
  • Use cases in mathematics and scientific problem-solving where logical execution is critical.
  • Developers looking for a compact model (2.6B parameters) with strong reasoning abilities for local or production deployment.