Moonlink/Mable-0.5

TEXT GENERATIONPricing:Input $0.32 / Cached $0.064 / Output $1.6Concurrent Unit Cost:1Model Size:2.6BQuant:BF16Context Size:8kPublished:Jul 22, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Moonlink/Mable-0.5 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, this model is optimized for tasks requiring step-by-step logical processing. It supports a context length of 8192 tokens and is available in GGUF, LoRA adapter, and merged 16-bit formats.

Loading preview...

Mable-0.5: A Reasoning-Focused Gemma-2-2B Variant

Mable-0.5, developed by Moonlink, is a 2.6 billion parameter language model built upon Google's Gemma-2-2B-it. This model is specifically fine-tuned for enhanced reasoning capabilities, distinguishing it from general-purpose LLMs.

Key Capabilities & Features

  • Specialized Reasoning: Excels in step-by-step reasoning, structured chain-of-thought (CoT) breakdown, and execution-oriented decision-making.
  • Fine-Tuning Data: Trained on 3,500 curated reasoning traces from the Fable-5 dataset, optimizing its logical processing.
  • Base Architecture: Utilizes the robust Gemma-2-2B-it architecture, enhanced through LoRA fine-tuning (Rank = 16, Alpha = 32) via Unsloth.
  • Flexible Deployment: Available in multiple formats including GGUF for local/Ollama/LM Studio, LoRA adapters for Transformers/Unsloth, and merged 16-bit weights for vLLM/production pipelines.
  • Gemma Chat Template: Follows the Gemma chat template, incorporating explicit THOUGHT: and ACTION: blocks for structured responses.

Ideal Use Cases

  • Complex Problem Solving: Suited for applications requiring detailed, multi-step logical deductions.
  • Automated Reasoning: Effective for tasks where a clear, traceable thought process is beneficial.
  • Structured Output Generation: Generates responses with explicit reasoning steps, making it valuable for explainable AI or agentic workflows.
  • Local Deployment: The GGUF format allows for efficient local execution on CPU or Apple Silicon, making it accessible for various development environments.

Finetunes
1 models