---

license: mit

base_model: Mia-AiLab/Qwable-3.6-27b

base_model_relation: quantized

pipeline_tag: text-generation

library_name: gguf

tags:

• gguf
• ollama
• imatrix
• qwen3
• qwable
• local-llm
• llama.cpp
• lm-studio
• quantized
• low-vram
• reasoning
• code

quantized_by: liodon-ai

---

Qwable-3.6-27b — iMatrix GGUF

The first iMatrix GGUF for Qwable-3.6-27b — Q2_K through Q8_0 with importance matrix calibration.

Qwable-3.6-27b is a fine-tune of Qwen 27B trained on Fable 5-style reasoning traces — it thinks before answering, with structured deliberate responses optimized for code and technical tasks.

These GGUFs are produced from the F16 source using importance matrix (iMatrix) calibration on 2M tokens of wikitext-103. iMatrix identifies which weights matter most during inference and protects them during quantization — the result is noticeably better coherence at Q2/Q3/Q4.

---

Quick Start

llama.cpp

llama-cli -hf liodon-ai/Qwable-3.6-27b-imatrix-GGUF:Q4_K_M

LM Studio / Jan

Search liodon-ai/Qwable-3.6-27b-imatrix-GGUF and pick your quant.

---

Available Quants

| Quant | Size | VRAM | Notes |

|-------|------|------|-------|

| Q2_K | 10.9 GB | 9 GB | tiniest — runs almost anywhere, iMatrix-improved |

| Q3_K_M | 13.5 GB | 11 GB | great for 8GB VRAM, iMatrix-improved |

| Q4_K_M | 16.8 GB | 14 GB | sweet spot (recommended), iMatrix-improved |

| Q5_K_M | 19.5 GB | 18 GB | high quality, iMatrix-improved |

| Q6_K | 22.4 GB | 20 GB | near-lossless, iMatrix-improved |

| Q8_0 | 29.0 GB | 28 GB | basically full quality |

---

Why iMatrix for Qwable?

Qwable uses chain-of-thought reasoning — it emits long <think> traces before answering. At low-bit quantization, coherence over long sequences matters more than for simple Q&A models. iMatrix protects the weights that sustain long reasoning chains, giving noticeably better output at Q2_K and Q3_K_M compared to standard quantization.

---

What is iMatrix?

Standard quantization rounds all weights equally. iMatrix:

1. Runs calibration text through the full-precision model
2. Measures which weights activate most (the "importance matrix")
3. Allocates more precision to important weights, less to unimportant ones

Same file size. Better output. Most noticeable at Q2/Q3/Q4.

---

Calibration

Importance matrix computed from 2M tokens of wikitext-103 — 128 calibration chunks.

---

Source Model

• Original: Mia-AiLab/Qwable-3.6-27b — 22.9K downloads
• Architecture: Qwen3.5 27B fine-tuned on Fable 5 reasoning traces
• Strengths: Code, debugging, technical reasoning, structured tasks
• License: MIT