Qwen3.8-27B-MagicQuant-GGUF

Qwen3.8-27B-MagicQuant-GGUF

Derivative of Qwen3.8-27B, quantized using MagicQuant hybrid evolutionary per-tensor search.

Sibling repo with AMD-native (ROCmFPX fork-only) builds: lmcoleman/Qwen3.8-27B-ROCmFPX-GGUF.

Base Model

This is a derivative of Qwen3.8-27B. 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)
• 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

Why there is no Q5

A Q5 tier was searched, built and measured. It is not published, because it was dominated by Q4 on every axis measured here:

It was 21% larger than Q4, measured slightly worse, and generated ~24% slower. The quality difference is small enough to be a tie rather than a real regression, but a tie at 21% more disk and a quarter less speed is not a tier worth shipping: there is no request for which it is the right answer.

This is a property of how the schemes round into size bands for this particular model, not a defect in the file. The useful ladder here is Q4 for speed, Q6 for quality.

Throughput figures above are CPU-only (llama-bench, no GPU offload, on a Ryzen AI MAX+ 395), measured during the search alongside other load. They are useful for comparing these tiers against each other -- generation rate was consistent across all 15 measured candidates -- but they are not the speed you should expect from a GPU or Metal build, and they are not a benchmark of this hardware. Prompt-processing figures from the same runs were inconsistent and are omitted for that reason.

Perplexity measured on wikitext-2 (100 chunks, ctx 512) against the BF16 baseline of 6.7443. 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.

Recommended: Q4 (14.65 GiB). It is the smallest tier that is statistically tied with the best measured quality here. Q6 is 43% larger for 0.048 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 Qwen3.8-27B-Q4_K_M.gguf -c 8192 --jinja -cnv

# Single prompt
llama-cli -m Qwen3.8-27B-Q4_K_M.gguf -c 8192 -p "Your prompt here"

# Server mode
llama-server -m Qwen3.8-27B-Q4_K_M.gguf -c 8192 --port 8080 --jinja

Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(model_path="./Qwen3.8-27B-Q4_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 Qwen3.8-27B-Q4_K_M.gguf --mmproj mmproj-Qwen3.8-27B-f16.gguf -c 8192 --port 8080 -ngl 99 -fa on

Serving: MTP Speculative Decoding

This model includes MTP ("nextn") draft tensors, enabling self-speculative decoding -- measured ~1.6-1.9x faster generation with a ~95% first-token accept rate (no separate draft model needed; it drafts from itself):

llama-server -m Qwen3.8-27B-Q4_K_M.gguf -c 8192 --port 8080 --host 127.0.0.1 -ngl 99 -md Qwen3.8-27B-Q4_K_M.gguf --spec-type draft-mtp -ctk q8_0 -ctv q8_0 -fa on

Memory cost: MTP needs its own draft context alongside the main context, so serving with it uses roughly 2x the model's memory compared to serving without -md/--spec-type draft-mtp.

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