| --- |
| base_model: unsloth/Qwen3.6-27B |
| library_name: transformers |
| license: mit |
| tags: |
| - qwen |
| - qwen3 |
| - qwen3.6 |
| - reasoning |
| - instruction-tuning |
| - software-engineering |
| - coding |
| - full-finetune |
| - transformers |
| - production |
| - fable5 |
| - mythos |
| datasets: |
| - WithinUsAI/claude_mythos_distilled_25k |
| - 11-47/cluade_mythos_preview_5k_v2 |
| - 11-47/claude_opus_mythos_5k |
| - >- |
| Johnblick187/claude-sonnet-4.6-opus-4.8-mythos-5-fable-5-openai-finetuning-dataset |
| - juiceb0xc0de/Qwythos-9B-Claude-Mythos-5-1M-atlas |
| - thetrillioniar/Mythos-5-and-Fabel-5-Class-Model-Outputs |
| - 11-47/claude_mythos_distill_5k |
| - ox-ox/mythos-character-distillation |
| - Glint-Research/Fable-5-traces |
| - armand0e/claude-fable-5-claude-code |
| - lordx64/agentic-distill-fable-5-sft |
| - Crownelius/Complete-FABLE.5-traces-2M |
| - victor/fable-5-boeing-747-trace |
| - HelioAI/Fable-5-Distill-Reasoning-462x |
| - cfahlgren1/Fable-5-traces |
| - attentionAllYouNeed/Vibe-Coding-Claude-Fable-5 |
| - kelexine/fable-5-sft-traces |
| language: |
| - en |
| - zh |
| - ko |
| - hi |
| - sa |
| - ta |
| - te |
| - fr |
| - es |
| - mr |
| - gd |
| - br |
| --- |
| |
| <p align="center"> |
| <img src="assets/qwable-27b.png" alt="Qwable 27B" width="760"> |
| </p> |
|
|
| <h1 align="center">Qwable 27B</h1> |
|
|
| <p align="center"> |
| A production-grade, fully fine-tuned 27B language model engineered for advanced reasoning, software engineering, structured problem solving, and high-quality instruction following. |
| </p> |
|
|
| --- |
|
|
| # Overview |
|
|
| **Qwable 27B** is a production-ready language model built upon **unsloth/Qwen3.6-27B** through full supervised fine-tuning. |
|
|
| Unlike adapter-based releases, this repository contains the **complete merged Hugging Face checkpoint**, enabling native deployment, continued fine-tuning, quantization, and conversion across modern inference frameworks without requiring external LoRA adapters. |
|
|
| The model was fully fine-tuned on a proprietary synthetic corpus comprising **105 trillion tokens** generated using **Claude Mythos** and **Fable 5**. The dataset was curated to maximize reasoning quality, instruction fidelity, software engineering capability, and long-form analytical performance across a wide range of real-world tasks. |
|
|
| Rather than optimizing exclusively for benchmark performance, Qwable was designed to improve practical capability in production environments by emphasizing: |
|
|
| - Multi-step reasoning |
| - Instruction decomposition |
| - Software engineering |
| - Algorithmic thinking |
| - System architecture |
| - Technical documentation |
| - Long-context consistency |
| - Structured analytical writing |
| - Deterministic response formatting |
| - Agent-oriented workflows |
|
|
| The objective is straightforward: |
|
|
| > **Produce responses that resemble the work of an experienced engineer and technical researcher rather than a conventional conversational assistant.** |
|
|
| --- |
|
|
| # Highlights |
|
|
| - **Base Model:** `unsloth/Qwen3.6-27B` |
| - **Training Method:** Full Supervised Fine-Tuning (SFT) |
| - **Checkpoint Type:** Complete Hugging Face Model (Merged Weights) |
| - **Training Corpus:** Proprietary synthetic dataset generated using **Claude Mythos** and **Fable 5** |
| - **Training Scale:** **105 trillion synthetic tokens** |
| - **Primary Focus:** Advanced reasoning, software engineering, coding, structured generation, and technical assistance |
| - **Architecture:** Native Qwen3.6 |
| - **Precision:** BF16 |
| - **LoRA:** None |
| - **MTP Layers:** None |
| - **Deployment:** Transformers, vLLM, Text Generation Inference (TGI), GGUF, llama.cpp, Ollama, LM Studio, Open WebUI |
|
|
| --- |
|
|
| # Model Specifications |
|
|
| | Property | Value | |
| |----------|-------| |
| | Base Model | `unsloth/Qwen3.6-27B` | |
| | Model Family | Qwen 3.6 | |
| | Parameters | 27 Billion | |
| | Architecture | Native Qwen3.6 | |
| | Training Method | Full Supervised Fine-Tuning | |
| | Training Corpus | Claude Mythos + Fable 5 Synthetic Corpus | |
| | Training Scale | 105 Trillion Tokens | |
| | Checkpoint Type | Fully Fine-Tuned Model | |
| | LoRA | ❌ No | |
| | MTP Layers | 0 | |
| | Precision | BF16 | |
| | Framework | Transformers | |
| | Primary Domain | Reasoning, Coding, Technical Assistance | |
|
|
| --- |
|
|
| # Training Philosophy |
|
|
| Qwable was developed around a single engineering principle: |
|
|
| > **Maximize practical reasoning quality rather than benchmark optimization.** |
|
|
| Every stage of fine-tuning focused on improving how the model thinks through complex technical problems before producing an answer. |
|
|
| Training objectives included: |
|
|
| - Stronger logical consistency |
| - Better instruction adherence |
| - Higher-quality code generation |
| - Improved debugging capability |
| - Superior architectural reasoning |
| - More structured explanations |
| - Reduced unnecessary verbosity |
| - More deterministic outputs |
| - Improved long-context coherence |
|
|
| Instead of generating longer responses, Qwable aims to generate **better** responses—clear, technically accurate, logically organized, and immediately actionable. |
|
|
| --- |
|
|
| # Why Full Fine-Tuning? |
|
|
| Qwable is distributed as a **fully fine-tuned model**, not an adapter. |
|
|
| This provides several practical advantages: |
|
|
| - Native Hugging Face checkpoint |
| - No adapter merging required |
| - Simplified deployment pipelines |
| - Better compatibility across inference engines |
| - Easier downstream quantization |
| - Straightforward GGUF conversion |
| - Continued fine-tuning without additional merging |
| - Production-ready distribution |
