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

Hugging Face
TEXT GENERATIONPricing:Input $0.468 / Output $1.82Concurrent Unit Cost:1Model Size:8BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jul 14, 2026License:apache-2.0Architecture:Transformer Open Weights Featherless Exclusive Warm

AnkitAI/Parable-Qwen3-8B-Claude-Fable-5 is an 8 billion parameter Qwen3 fine-tune, developed by AnkitAI, specifically trained on real multi-step agent traces from Claude Fable 5 and GPT-5.5. This model demonstrates a 67% lower held-out test loss compared to its base and achieves the highest strict qualitative score in the Parable series for agent work. It is optimized for complex reasoning, planning, and tool use, making it suitable for agentic applications.

Loading preview...

Overview

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

Key Capabilities & Performance

  • Agentic Reasoning: Fine-tuned specifically for agent work, excelling in tasks requiring multi-step planning and tool utilization.
  • Improved Performance: Achieves a 67% lower held-out test loss (0.712) compared to the base Qwen3-8B (2.162).
  • High Qualitative Score: Demonstrated the strongest strictly-graded qualitative score in the Parable series, with 23 out of 34 prompts fully correct and 30/34 correct or partially correct in coding, terminal, and debugging tasks.
  • Context Length: Inherits the base model's 128K-token context, with fine-tuned behavior strongest in the opening turns of a session (trained at 1,024-token sequences).

Training Data & Methodology

The model was fine-tuned using QLoRA on a single 16 GB GPU, utilizing datasets such as Glint-Research/Fable-5-traces (4.4k Claude Fable 5 coding-agent session traces) and Roman1111111/gpt5.5-terminal (terminal-agent task solutions).

Limitations

  • Function Calling Trade-off: While optimized for reasoning, it shows a slight trade-off in function-calling accuracy (2.3 to 5.0 points lower on BFCL V3 categories compared to the base Qwen3-8B).
  • Agent-Oriented Responses: May sometimes respond with structured tool-call JSON instead of prose for ops-style prompts, which is useful within agent harnesses but might require re-prompting in plain chat.

Finetunes
2 models