gustajunq/openFable-4B-general

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jun 25, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

OpenFable-4B is a 4 billion parameter language model developed by gustajunq at SynastrIA Networks, fine-tuned from Qwen3-4B with a 32768 token context length. It is specifically designed to replicate the conversational style, reasoning depth, and structured output quality of Claude Fable 5. This model excels in direct, warm, and structured responses across coding, math, agentic planning, and cybersecurity tasks, making it suitable for applications requiring a distinct conversational persona.

Loading preview...

OpenFable-4B: A Style-First Conversational Model

OpenFable-4B, developed by gustajunq at SynastrIA Networks, is a 4 billion parameter model fine-tuned from Qwen3-4B. Its primary distinction lies in its "style-first" approach, meticulously crafted to emulate the conversational tone, reasoning capabilities, and structured output characteristic of Claude Fable 5. This is achieved through a custom-built dataset of approximately 300 hand-curated examples spanning coding, mathematics, agentic planning, and cybersecurity, rather than relying on generic synthetic datasets.

Key Capabilities & Differentiators

  • Distinct Conversational Style: Engineered for direct, warm, structured, and non-verbose responses, with a custom chat template embedding a unique personality.
  • Custom Dataset: Utilizes a proprietary dataset focused on technical and reasoning tasks, avoiding common CoT preambles.
  • Strong Reasoning: Despite being a style fine-tune, it achieves a 68.48% MMLU score (zero-shot) and matches top-tier 4B models on GSM8K math reasoning, closing the gap with its Qwen3-4B base.
  • Local Inference Ready: Provided in Q4_K_M GGUF quantization, optimized for local deployment via llama.cpp, LM Studio, and similar platforms.

Ideal Use Cases

  • Applications requiring a specific, consistent conversational persona akin to Claude Fable 5.
  • Tasks involving structured output in technical domains like coding, math, and cybersecurity.
  • Edge devices or local environments benefiting from efficient 4B parameter models with a 32768 token context length.

Limitations

  • Performance in Humanities is lower (59.5% MMLU) due to dataset focus.
  • Primarily optimized for English, not multilingual use.

Finetunes
2 models