Parable models are 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.

This page hosts the Qwen3 line of the family. It adds structured <think> reasoning and agent-style task execution to Qwen3, tuned for local coding and terminal work.

Highlights:

Trained on real agent traces: multi-step tool use, planning, and <think> reasoning from actual Claude Fable 5 and GPT-5.5 sessions, not synthetic Q&A.
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.
Licence: Apache-2.0 weights, ready for local and commercial use (see data provenance note below).

Evaluation (held-out test split, identical eval code for base and fine-tune):

4b test loss: 1.888 (base) to 0.996, a 47% reduction
4b token accuracy: 0.683 (base) to 0.782

Token accuracy 4b: 0.683 → 0.782. Strictly graded qualitative results are published on the Hugging Face cards.

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.

Family

Granite line: parable/granite4.1-fable
Full-precision weights, quant tables, and complete eval reports: Hugging Face, AnkitAI

Get Parable

LM Studio: this page (lms get)
Ollama: ollama run parable/qwen3-fable, plus the sibling line and the family flagship parable/fable
Hugging Face: full weights, GGUF quants, eval reports, datasets
LM Studio: search "parable", or paste any Parable Hugging Face URL

Base model: Qwen3 (Apache-2.0). 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.

This page hosts the Qwen3 line of the family. It adds structured <think> reasoning and agent-style task execution to Qwen3, tuned for local coding and terminal work.

Highlights:

Trained on real agent traces: multi-step tool use, planning, and <think> reasoning from actual Claude Fable 5 and GPT-5.5 sessions, not synthetic Q&A.
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.
Licence: Apache-2.0 weights, ready for local and commercial use (see data provenance note below).

Evaluation (held-out test split, identical eval code for base and fine-tune):

4b test loss: 1.888 (base) to 0.996, a 47% reduction
4b token accuracy: 0.683 (base) to 0.782

Token accuracy 4b: 0.683 → 0.782. Strictly graded qualitative results are published on the Hugging Face cards.

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.

Family

Granite line: parable/granite4.1-fable
Full-precision weights, quant tables, and complete eval reports: Hugging Face, AnkitAI

Get Parable

LM Studio: this page (lms get)
Ollama: ollama run parable/qwen3-fable, plus the sibling line and the family flagship parable/fable
Hugging Face: full weights, GGUF quants, eval reports, datasets
LM Studio: search "parable", or paste any Parable Hugging Face URL

Base model: Qwen3 (Apache-2.0). 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.

qwen3-fable

qwen3-fable

Get Parable

Get Parable