Subscribe to my newsletter · Support

Qwen3.8-27B-SLQ-GGUF

This is Qwen/Qwen3.8-27B, compressed with a customized version of SLQ (Statistically-Lossless Quantization) and assembled as GGUF. The search combines existing quantized donor weights and chooses a source separately for 498 complete matrices, including embeddings and the output head.

Choose one of three models:

• EAR targets: 98%, 98.5%, and 99%, selected using the BF16 reference's top 10 tokens.
• Sizes: decimal GB (1 GB = 1,000,000,000 bytes). Active text weights include fixed text tensors and exclude MTP, metadata and alignment.
• Embeddings / output head: Q4_K / Q8_0 for EAR98; Q4_K / BF16 for EAR98.5; Q6_K / BF16 for EAR99.
• Each file retains 351,008,768 bytes of quantized MTP tensors from the UD-Q5_K_XL donor. MTP was disabled during evaluation. No vision tensors or multimodal projector are included.
• Made by: The Kaitchup.
• License: Apache 2.0.

The filename suffixes Q5_K_S, Q5_K_M, and Q6_K are Hub compatibility and dominant-precision-family labels for these custom SLQ allocations.

Method: SLQ paper · SLQ repository.

Candidate sources

The search considered UD-Q4_K_XL, UD-Q5_K_XL, UD-Q6_K_XL, Q4_0, Q8_0, and BF16 from unsloth/Qwen3.8-27B-GGUF, plus IQ3_S from ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF. Identical tensor payload alternatives were canonicalized. The final three models select no GSQ donor tensors; its candidates remained eligible during search.

Credit goes to the Qwen team for the base model, Unsloth for the GGUF donors, ISTA-DASLab for GSQ/RCO, and Michael Helcig, Eldar Kurtic and Dan Alistarh for SLQ. The base and donor repositories declare Apache 2.0 licensing; the full license is included below. Original copyright and attribution rights remain with their authors.

Precision by layer

EAR98

SVG

EAR98.5

SVG

EAR99

SVG

EAR comparison

EAR means Expected Acceptance Rate. It measures the overlap between the original BF16 model's next-token probabilities and those of the compared model. For EAR@K, both distributions are separately renormalized over the BF16 reference's K most likely tokens. Full-vocabulary EAR uses all 248,320 token IDs. An EAR of 0.99 means 99% average probability overlap on that measured support; it does not mean 99% benchmark accuracy or identical generated answers.

Zoomable SVG

SLQ's “statistically lossless” objective concerns approximate fidelity, not identical numerical outputs. These releases meet their stated measured EAR@10 targets. The same calibration corpus guided model selection, so these values are calibration fidelity rather than an independent held-out benchmark. Statistical equivalence on downstream tasks has not been established for these releases.

Instructions: llama.cpp

Use a llama.cpp build with Qwen3.8 architecture support. Native text scoring and standalone GGUF parity were validated on an NVIDIA GH200 with llama.cpp commit 45451c366d4256f94fbc8845645e31a6b9155054. This validation did not cover image/video inference or speculative decoding.

Install the Hugging Face CLI, then download one file:

pip install -U huggingface_hub
hf download kaitchup/Qwen3.8-27B-SLQ-GGUF \
  Qwen3.8-27B-SLQ-EAR98-Q5_K_S.gguf --local-dir .

Start an OpenAI-compatible text server with the downloaded model:

llama-server \
  --model Qwen3.8-27B-SLQ-EAR98-Q5_K_S.gguf \
  --n-gpu-layers 99 \
  --ctx-size 8192 \
  --port 8080

To use EAR98.5 or EAR99, download the corresponding file from the table and change --model.

Send a chat request:

curl http://localhost:8080/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "messages": [{"role": "user", "content": "Is Earth smaller than Venus?"}],
    "max_tokens": 256
  }'

Support

If these models are useful, support the work or subscribe to The Kaitchup.