AnkitAI/Parable-Qwen3-4B-Claude-Fable-5
AnkitAI/Parable-Qwen3-4B-Claude-Fable-5 is a 4 billion parameter Qwen3-based model fine-tuned for local coding and agent-shaped reasoning. It distills planning, tool habits, and terminal reasoning from genuine Claude Fable 5 agent sessions, not synthetic Q&A. This model excels at providing structured, reliable answers for multi-step agent workflows and coding tasks, even on modest hardware. It offers improved answer rates and agent-trace loss compared to its base model, making it suitable for offline agent and coding work.
Loading preview...
Parable-Qwen3-4B-Claude-Fable-5: A Local Coding Agent Model
AnkitAI/Parable-Qwen3-4B-Claude-Fable-5 is a 4 billion parameter model built on the Qwen3-4B base, specifically designed for local coding and agent-like reasoning. Its unique training methodology involves distilling planning, tool usage, and terminal reasoning directly from real Claude Fable 5 agent sessions, rather than relying on synthetic question-and-answer data. This approach imbues the model with structured, agent-shaped reasoning capabilities.
Key Capabilities and Features
- Reliable Answering: Unlike its base model, which can fail to answer 34% of prompts, this model consistently provides responses, significantly reducing empty outputs.
- Agent-Shaped Reasoning: Trained on genuine multi-step agent sessions, it produces structured plans, tool selections, and terminal workflows, making its reasoning more organized and less improvised.
- Lightweight and Local: With 4 billion parameters and a GGUF build of approximately 2.5 GB, it is designed to run efficiently on laptops, older GPUs, or modest desktops, keeping code and data local.
- Improved Performance (v2.1): The latest recalibrated merge (v2.1) shows a +1.8 point improvement on HumanEval-164, reaching 74.4, and significantly lower held-out agent-trace loss (1.876 vs 2.846) compared to the base Qwen3-4B.
When to Use This Model
This model is particularly well-suited for:
- Local Agent and Coding Work: Ideal for scenarios where structured, reliable answers are paramount, especially in multi-step agent workflows.
- Offline Development: Its small footprint allows for offline execution, ensuring data privacy and accessibility without cloud dependencies.
- Structured Reasoning Tasks: When the style of reasoning, including planning and tool use, is as important as the final output, this model's agent-trace training provides an advantage.
For maximum-accuracy function calling in a tool-calling harness where reasoning style is less critical, the base Qwen3-4B might offer a slight edge in specific benchmarks, but this model provides a more robust and consistent agent experience.