AnkitAI/Parable-Qwen3-4B-Claude-Fable-5

TEXT GENERATIONConcurrent Unit Cost:1Model Size:4BQuant:BF16Context Size:32kTool Calling:SupportedPublished:Jul 9, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

AnkitAI/Parable-Qwen3-4B-Claude-Fable-5 is a 4 billion parameter Qwen3-4B fine-tuned model developed by AnkitAI, specifically optimized for agentic reasoning and multi-step tool use. It was trained on real Claude Fable 5 and GPT-5.5 agent traces, focusing on planning and reasoning captured from actual agent sessions. This model achieves a 0.782 token accuracy on held-out test data, demonstrating strong performance in agentic coding behaviors with a 32768 token context length.

Loading preview...

Parable-Qwen3-4B-Claude-Fable-5 Overview

AnkitAI/Parable-Qwen3-4B-Claude-Fable-5 is the inaugural release in the Parable series, a 4-billion parameter model based on Qwen3-4B. This model is uniquely fine-tuned on authentic agent behavior traces from Claude Fable 5 and GPT-5.5, rather than synthetic Q&A. Its training emphasizes multi-step tool use, complex planning, and detailed <think> reasoning processes, making it particularly adept at simulating agentic workflows.

Key Capabilities & Performance

  • Agentic Reasoning: Excels in tasks requiring multi-step planning and tool utilization, derived from real-world agent sessions.
  • High Accuracy: Achieves a token accuracy of 0.782 on held-out test data, representing a 47% reduction in test loss compared to the base Qwen3-4B model.
  • Coding & Debugging: In qualitative reviews, 92% of prompts that produced a final answer were correct for coding, terminal, and debugging tasks.
  • Reasoning Output: Generates a detailed <think>...</think> block before the final answer, providing insight into its problem-solving process.

Ideal Use Cases

  • Agentic Workflows: Suited for applications requiring models to perform multi-step tasks, planning, and tool invocation.
  • Code Generation & Debugging: Effective for generating Python functions, solving terminal tasks, and assisting with debugging.
  • Reasoning-Intensive Applications: Best utilized where the model's internal thought process (via the <think> block) is valuable or can be leveraged.

Important Considerations

  • Token Budget: Requires a generous max_new_tokens (recommended ≥ 2500) as it invests heavily in its reasoning phase. Insufficient budget can lead to incomplete responses.
  • Specialized Focus: Tuned specifically for agentic coding behavior, which may trade some general knowledge breadth for specialized depth.
  • Output Verification: As with many small models, critical outputs like generated commands and code should be reviewed and verified.

Finetunes
2 models