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

TEXT GENERATIONPricing:Input $0.334 / Output $1.71Concurrent 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 a fine-tuned Qwen3-8B model, developed by AnkitAI, specifically optimized for multi-step agent work. This 8 billion parameter model was trained on real Claude Fable 5 and GPT-5.5 agent traces, including planning, tool use, and reasoning. It achieves a 67% lower held-out test loss compared to its base model and demonstrates strong qualitative performance in coding, terminal, and debugging tasks, making it suitable for complex agentic applications.

Loading preview...

Model Overview

AnkitAI/Parable-Qwen3-8B-Claude-Fable-5 is a specialized fine-tune of the Qwen3-8B base model, developed by AnkitAI. Its core distinction lies in its training data: real multi-step agent sessions from Claude Fable 5 and GPT-5.5, encompassing planning, tool use, and explicit <think> reasoning. This approach aims to imbue the model with advanced agentic capabilities, moving beyond synthetic Q&A.

Key Capabilities & Performance

  • Agentic Reasoning: Trained on actual agent traces, it excels in tasks requiring multi-step planning and tool utilization.
  • Reduced Test Loss: Achieves a significant 67% lower held-out test loss compared to the base Qwen3-8B model.
  • Strong Qualitative Score: In a strict qualitative review of 34 coding, terminal, and debugging prompts, it scored 23/34 fully correct and 30/34 correct or partially correct, marking the highest fully-correct score in the Parable series.
  • Efficient Fine-tuning: Fine-tuned using QLoRA (NF4, 1024 sequence length) on a single 16 GB GPU.

Limitations & Considerations

  • Agent-centric Output: When used outside agent harnesses, it may occasionally respond with structured tool-call JSON instead of prose (2/34 times in evaluations).
  • Function Calling Trade-off: While strong in reasoning, it shows a slight trade-off in function calling accuracy (2.3 to 5.0 points lower on BFCL V3) compared to the base Qwen3-8B.
  • Context Length: Fine-tuned at 1,024-token sequences, though the base model's native 128K-token context remains available for longer sessions, with fine-tuned behavior strongest in initial turns.

Ideal Use Cases

This model is particularly well-suited for applications requiring advanced agentic behavior, complex problem-solving, and multi-step reasoning, especially in coding, debugging, and terminal-based tasks. Developers building agent harnesses or systems that benefit from explicit reasoning traces will find this model highly effective.

Finetunes
2 models