mfielding92

SmartCode-Fable-5-CoT-Reasoning-QVO-Qwen-3.6-27B-Distilled

README

License: apache-2.0

🎩 Layer 1: The ThinkingCap Foundation

The base model isn't just another Qwen3.6-27B repack. ThinkingCap is the finetune that took a serious swing at reasoning bloat:

⚡ ~50% fewer thinking tokens on average than stock Qwen3.6-27B.
🔥 Over 90% reduction in best-case scenarios. It stops re-deriving the obvious.
🎯 Less than 1% average accuracy difference across benchmarks compared to the original base.

Translation: you keep the intelligence, you drop the rambling. No more 3,000 tokens of "wait, but actually, let me reconsider" before printing a two-line function.

🍜 Layer 2: The SmartCode Fable 5 Injection

On TOP of that efficiency, this finetune pours in 1,000 curated reasoning sequences distilled from Fable 5, and with them, a genuine coding personality:

🗣️ It talks to itself as it codes. The model narrates its work in real time, catching mistakes before they land in the output.
🗺️ It plans out loud before writing a line. What it's building, how, and why, the same discipline agentic coding tools enforce, baked directly into the weights.
📈 Adaptive Reasoning depth. Chain-of-Thought scales with task complexity. Easy problems get short chains. Hard problems get the full Fable 5 treatment.
🧬 No complexity overload. Most distills fail by cramming frontier-sized reasoning into a model that can't carry it. This dataset was built to fit.

Translation: it doesn't just answer, it works the problem. Plan first, code second, accuracy up.

The two layers fit together well. ThinkingCap was already trained to think only as much as necessary, and Adaptive Reasoning data teaches exactly that behavior at the frontier level. The techniques reinforce each other instead of fighting. You get frontier-flavored coding reasoning that terminates when it's actually done thinking.

🎯 READ THIS: Recommended Sampler Settings

This model was tuned around these exact settings. Do not skip this section. Bad sampler settings are the #1 cause of "the distill is broken" reports.

Table with columns: Setting, Value
Setting	Value
Temperature	0.9
Top P	0.95
Top K	60
Min P	0
Repeat Penalty	NONE (1.0 / disabled)
Presence Penalty	1.0

⚠️ No repetition penalty by default. It mangles code formatting. Only raise repeat or presence penalty if you actually see looping, and raise it as little as possible.

Architecture: qwen3_5 (27B parameters)
Base: bottlecapai/ThinkingCap-Qwen3.6-27B
Max sequence length: 8,192 tokens per training entry
Reasoning cap in training data: under 2,048 tokens, for density and balance
Training samples: 1,000 Fable 5 distilled CoT sequences
Trained 2x faster with Unsloth and Huggingface's TRL library

Standard Qwen3.6 chat template applies. Keep the reasoning block enabled, since this model's entire value proposition lives in how efficiently it uses that block.

Because both the base finetune and this distill were tuned toward short, decisive reasoning, aggressive thinking-budget forcing is generally unnecessary and may hurt output quality on the hard tail of problems. Let it decide.

The model weights are released under apache-2.0.

However, the SmartCode Fable 5 Distill dataset it was trained on is designated personal hobbyist and educational use only, with commercial use prohibited to remain in alignment with provider policies. If you care about staying clean on distillation-provenance grounds, treat this model the same way: personal and educational use.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

mfielding92

Model Tree

Base

bottlecapai/ThinkingCap-Qwen3.6-27B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality