Qwen3.6-27B-Fable-Fusion-711-MTP-MagicQuant-GGUF

Derivative of Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP, quantized using MagicQuant hybrid evolutionary per-tensor search.

Sibling repo with AMD-native (ROCmFPX fork-only) builds: lmcoleman/Qwen3.6-27B-Fable-Fusion-711-MTP-ROCmFPX-GGUF.

Base Model

This is a derivative of Qwen3.6-27B-Fable-Fusion-711-Uncensored-Heretic-NM-DAU-MTP. All credit for the base model architecture and weights goes to the original authors. The base model's license applies to this derivative.

Quantization Method

Quantized using MagicQuant hybrid evolutionary per-tensor quantization, based on the methodology by magiccodingman:

• Tensors are classified into sensitivity groups (Embeddings, Head, Query, Key, Output, FFN Up/Down, MoE Experts, Router)
• An evolutionary search finds the optimal quantization type per group, balancing size vs. perplexity
• Q4/Q5/Q6 tier targets are searched, and each one ships only if it earns its place (see below); a Q8-band build was added 2026-08-08 from a re-run of the search
• 2026-08-08: the search was re-run with KL-resolved sensitivity probing (the original probe metric could not resolve sub-percent group deltas on a dense model). The re-run re-validated the Q4/Q5/Q6 configs unchanged -- the published files remain the measured band winners -- and found the new Q8-band winner below
• Small-row tensors and sensitivity-critical layers (embeddings, output head, router) are kept at F32/F16/BF16
• This is NOT a uniform quantization -- each tensor group gets its own optimal type

A tier name here is a size band, not a promise that every tensor uses that exact type. A "Q5" is whatever mix of schemes landed in the Q5 size band with the lowest measured perplexity loss -- which is the point of the search.

GGUF Files

Perplexity measured on wikitext-2 (100 chunks, ctx 512) against the BF16 baseline of 6.2356. Lower is better; the percentage is the increase over BF16. These are the same measurements the tier selection is based on, so a tier that shipped is one that earned its size.

* The Q8's number is a 100-chunk spot-check of the freshly rendered file; in the re-run's own same-conditions measurement the Q8 and Q6 were statistically tied (difference far below measurement noise). Both readings say the same thing: on wikitext this Q8 is not measurably better than Q6.

About the Q8 (added 2026-08-08)

Config: FFN down/up at Q6_K, query at Q5_K, the SSM group at Q8_0 (its f32-required state operands fall back to F32 automatically), and everything else -- embeddings, head, key, output -- at BF16 (written as F16 on disk, a llama.cpp compute-graph limitation). It exists for users who want the "brains" of the model at effectively full precision and have the memory to spare; if you want the best quality per GB, Q6 (or the recommended Q5) is the better pick. It replaces a 31.3 GB Q8-band config found by the earlier, partially probe-blind search, at 5+ GB smaller and equivalent measured quality. Per this repo's tier semantics, "Q8" is a size band -- this file contains zero Q8_0-quantized attention tensors, and that is by measurement, not accident.

Recommended: Q5 (17.68 GiB). It is the smallest tier that is statistically tied with the best measured quality here. Q6 is 18% larger for 0.019 percentage points of perplexity, which is below what this measurement can resolve -- so the extra bytes buy nothing you can detect.

Usage

LM Studio

1. Download the GGUF file of your preferred quantization tier
2. Place it in your LM Studio models directory
3. Load the model in LM Studio -- it will auto-detect the chat template
4. The model supports the base model's full context length

llama.cpp

# Interactive chat (--jinja uses the model's embedded chat template, not a hardcoded one)
llama-cli -m Qwen3.6-27B-Fable-Fusion-711-MTP-Q5_K_M.gguf -c 8192 --jinja -cnv

# Single prompt
llama-cli -m Qwen3.6-27B-Fable-Fusion-711-MTP-Q5_K_M.gguf -c 8192 -p "Your prompt here"

# Server mode
llama-server -m Qwen3.6-27B-Fable-Fusion-711-MTP-Q5_K_M.gguf -c 8192 --port 8080 --jinja

Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(model_path="./Qwen3.6-27B-Fable-Fusion-711-MTP-Q5_K_M.gguf", n_ctx=8192)
output = llm.create_chat_completion(
    messages=[
        {"role": "user", "content": "Hello, how are you?"}
    ]
)
print(output["choices"][0]["message"]["content"])

Vision (image input)

llama-server -m Qwen3.6-27B-Fable-Fusion-711-MTP-Q5_K_M.gguf --mmproj mmproj-F16.gguf -c 8192 --port 8080 -ngl 99 -fa on

Caveats

• The base model's license (apache-2.0) applies to all derivative files
• Quantization reduces precision -- verify outputs for your specific use case
• The hybrid quantization assigns different precision to different tensor groups, which means quality characteristics may differ from uniform quantizations

Limitations

• Quantized models may exhibit subtle differences from the full-precision fine-tune
• This model inherits any limitations and biases present in the base model

Generated with MagicQuant