Qwen3.8 27B GSQ RCO GGUF

Non-uniform GGUF quantizations produced with GSQ and RCO, with a vision projector for multimodal use.

Overview

This repository provides GGUF quantizations of Qwen3.8-27B at four sizes, together with the model's vision projector (mmproj) for multimodal use. In contrast to uniform quantization, which applies a single quantization type to all weight tensors, each model here assigns a separate quantization type to every tensor. The assignment is obtained by a gradient-based search that allocates precision according to per-tensor sensitivity, subject to a total size budget. The resulting files are standard GGUF and run unmodified in llama.cpp, Ollama, and LM Studio.

Both methods were developed at the Deep Algorithms and Systems Lab (DASLab), Institute of Science and Technology Austria.

Available files

Files follow the convention <model>-GSQ-RCO-<type>.gguf, where the suffix names the quantization class; the table lists each file's true whole-file average bit-width. The mmproj file carries the vision encoder and projector at BF16; one copy serves all quantizations.

Each quantization also ships an optional -mtp build (about 0.35 GB larger) that carries the model's Multi-Token Prediction head for speculative decoding in llama.cpp. The weights are otherwise identical, so quality is unchanged.

The IQ3_S model is the task-lossless operating point: it matches the base model exactly on AIME25 (100.00) and LiveCodeBench v6 (85.71) and is within 0.51 points on GPQA-Diamond, at just over one fifth of the BF16 size.

Results

All models are evaluated against the BF16 base model and the Unsloth Dynamic (UD) quantizations of the same base model. We report perplexity on wikitext2, C4, and FineWeb-Edu, the average over five zero-shot tasks (arc_easy, arc_challenge, hellaswag, winogrande, piqa), recovery (zero-shot average relative to BF16), and three reasoning and generation benchmarks: AIME25, GPQA-Diamond, and LiveCodeBench v6. Sizes are those of the files as evaluated.

At 3.50 bpw, IQ3_S is task-lossless: it reproduces the base model exactly on AIME25 (100.00) and LiveCodeBench v6 (85.71) and trails it by 0.51 points on GPQA-Diamond, giving a task average of 91.70 against the base model's 91.87 (99.8%) at 11.8 GB, a 4.6x size reduction. Against UD-IQ3_S it leads by 3.33 points on AIME25 and 1.71 on LiveCodeBench while being 0.2 GB smaller, though UD holds GPQA-Diamond by 0.51. At 3.00 bpw the model already matches the base on AIME25 at 10.1 GB, and at matched file size (8.4 GB) IQ2_XS leads UD-IQ2_S by 10.00 points on AIME25, 8.59 on GPQA-Diamond, and 4.57 on LiveCodeBench v6.

Usage

llama.cpp

hf download ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF Qwen3.8-27B-GSQ-RCO-IQ3_XXS.gguf --local-dir .

llama-cli -m Qwen3.8-27B-GSQ-RCO-IQ3_XXS.gguf -p "Explain mixed-precision quantization." -ngl 99

Vision (multimodal)

hf download ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF mmproj-Qwen3.8-27B-BF16.gguf --local-dir .

llama-mtmd-cli -m Qwen3.8-27B-GSQ-RCO-IQ3_XXS.gguf \
  --mmproj mmproj-Qwen3.8-27B-BF16.gguf \
  --image photo.jpg -p "Describe this image."

The projector was converted directly from the base checkpoint and verified against these quantizations.

Ollama

ollama run hf.co/ISTA-DASLab/Qwen3.8-27B-GSQ-RCO-GGUF   # pick the file matching your memory budget

LM Studio

Search the repo name, then pick a GSQ-RCO-* build from the file list.

Quantization procedure

1. Per-tensor database. Each weight tensor is quantized at every candidate GGUF quantization type with GSQ, yielding a searchable database of quantized tensor variants.
2. RCO search. The budget-constrained Riemannian search assigns one quantization type per tensor such that the whole-file average bit-width meets the target.
3. Assembly. The selected per-tensor variants are stitched into a single standard GGUF file.

Reference implementations: GSQ at IST-DASLab/GSQ and RCO at IST-DASLab/RCO.

Reproducibility artifacts

Each released GGUF ships the files needed to audit how it was built:

The -mtp builds have their own allocation dumps; they list the same per-tensor assignment as the base model plus the 15 tensors of the MTP head.

Citation

If you use these models or methods, please cite both papers:

@article{gsq2026,
  title  = {GSQ: Highly-Accurate Low-Precision Scalar Quantization for LLMs via Gumbel-Softmax Sampling},
  author = {Dadgarnia, Alireza and Tabesh, Soroush and Nikdan, Mahdi and Helcig, Michael and Kurtic, Eldar and Kleinegger, Maximilian and Alistarh, Dan},
  journal= {arXiv preprint arXiv:2604.18556},
  year   = {2026}
}
@article{rco2026,
  title  = {Model Compression with Exact Budget Constraints via Riemannian Manifolds},
  author = {Helcig, Michael and Alistarh, Dan},
  journal= {arXiv preprint arXiv:2605.00649},
  year   = {2026}
}

Acknowledgements

We thank Verda and Scientific Computing at the Institute of Science and Technology Austria for providing the compute resources used to produce these models.

License

These quantized weights inherit the license of the base model (Qwen3.8-27B). The GSQ-RCO tooling is released by the Deep Algorithms and Systems Lab under its repository license.