AnkitAI/Parable-Qwen3-8B-Claude-Fable-5

TEXT GENERATIONConcurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 14, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Cold

AnkitAI/Parable-Qwen3-8B-Claude-Fable-5 is an 8 billion parameter Qwen3-8B fine-tune, developed by AnkitAI, specifically trained on real multi-step agent sessions from Claude Fable 5 and GPT-5.5. This model excels at agent work, including planning, tool use, and structured reasoning, demonstrating a 67% lower held-out test loss compared to its base. It is optimized for complex coding, terminal, and debugging prompts, making it suitable for agentic applications requiring structured thought processes.

Loading preview...

Model Overview

AnkitAI/Parable-Qwen3-8B-Claude-Fable-5 is an 8 billion parameter fine-tuned model based on Qwen/Qwen3-8B. It is distinguished by its training on real multi-step agent sessions, specifically capturing planning, tool use, and <think> reasoning from Claude Fable 5 and GPT-5.5 agent work, rather than synthetic Q&A.

Key Capabilities

  • Agentic Reasoning: Optimized for complex agent work, including multi-step planning and tool utilization.
  • Performance: Achieves a 67% lower held-out test loss compared to the base Qwen3-8B model.
  • Qualitative Accuracy: Demonstrates strong performance on strictly graded coding, terminal, and debugging prompts, with 23 out of 34 fully correct responses in internal evaluations.
  • Structured Output: Can generate structured tool-call JSON, which is beneficial for integration into agent harnesses.

Training Details

The model was fine-tuned using QLoRA (NF4, sequence length 1024) on a single 16 GB GPU. The training data includes:

Limitations

  • Primarily trained for agent work; may sometimes respond with structured tool-call JSON instead of prose in plain chat scenarios.
  • Fine-tuned at 1,024-token sequences, though the base model's 128K-token context is available, with fine-tuned behavior strongest in initial turns.

Good For

  • Developing AI agents that require multi-step reasoning, planning, and tool use.
  • Applications involving complex coding, debugging, and terminal-based tasks.
  • Scenarios where structured thought processes and outputs are beneficial.

Finetunes
2 models