Parable is a family of open models trained on real agent work: multi-step tool use, planning, and thinking traces captured from Claude Fable 5 and GPT-5.5 sessions, not synthetic Q&A. Every release is eval-gated against its base before it ships.

Highlights:

• Trained on real agent traces: tool use, planning, and <think> reasoning from actual Claude Fable 5 and GPT-5.5 sessions.
• Eval-gated releases: every model must beat its base on a held-out test split before it ships. Strictly graded qualitative results are published, which is rare in this niche.
• Best model per size: each size tag on this page carries the strongest Parable at that size. Base lines also ship on their own pages: parable/qwen3-fable, parable/granite4.1-fable.
• Licence: Apache-2.0 weights, ready for local and commercial use.

3B

ollama run parable/fable:3b

IBM Granite 4.1 3B line, chat-tuned on the prose side of the traces. Test loss 2.824 (base) to 0.376, an 87% reduction. Strictly graded qualitative: 14 of 34 fully correct, 26 of 34 correct or partially correct. Strongest at explanations, one-liners, and idiomatic refactors; for multi-part script generation use the 8B.

4B

ollama run parable/fable:4b

Qwen3-4B line. Test loss 1.888 (base) to 0.996, a 47% reduction. Token accuracy 0.683 to 0.782.

8B

ollama run parable/fable:8b

IBM Granite 4.1 8B line. Test loss 2.030 (base) to 0.617, a 70% reduction. Strictly graded qualitative: 20 of 34 fully correct, 32 of 34 correct or partial.

Usage notes

• Responses open with a <think>...</think> block before the final answer.
• Budget at least 2500 tokens for generation; reasoning models think before they answer.
• Sampling: temperature 0.7, top_p 0.95.

Get Parable

• Ollama: this page, plus parable/qwen3-fable and parable/granite4.1-fable
• Hugging Face: full weights, GGUF quants, eval reports, datasets
• LM Studio: search “parable”, or paste any Parable Hugging Face URL

References

• Full eval reports and full-precision weights on Hugging Face
• Fable-5 trace dataset in ChatML format

Training data: Glint-Research/Fable-5-traces (AGPL-3.0), gpt5.5-terminal (MIT). Third-party assistant traces; providers’ terms may apply to downstream training.