A ~16 GB imatrix-calibrated, mixed-precision GGUF of Qwen3.8-27B (a qwen35 hybrid SSM+attention architecture) — SGLabs' first dense-model Pym quant.

Base model: Qwen/Qwen3.8-27B — a multimodal (image-text-to-text), 27B dense model with 256K context, agentic-coding and reasoning strengths, and a native MTP head.

This quant: dense FFN to 2–3 bits; attention, SSM, norms, embeddings, and the MTP head preserved at Q8 (so speculative decoding stays on). Runs on llama.cpp — AMD (ROCm) and Apple Silicon. No CUDA required.

Precision map

ffn_gate, ffn_up → IQ2_XXS
ffn_down → IQ3_XXS
attention · SSM · norms · embeddings · output · MTP block → Q8_0

Usage

bash

# with MTP speculative decoding (recommended):
llama-server -m Qwen3.8-27B-Pym-IQ2_XXS.gguf -ngl 999 -c 32768 \
  --spec-type draft-mtp --spec-draft-n-max 2 -np 1

File / compatibility


File	`Qwen3.8-27B-Pym-IQ2_XXS.gguf`
Size	15.99 GB (14.9 GiB) · 4.68 BPW
Architecture	`qwen35` · 65 blocks (64 + MTP)

Mixed-precision quant: the filename carries the dominant standard token (IQ2_XXS) for general compatibility; loaders requiring a uniform general.file_type may report it as custom.

Quantized by SGLabs.

SGLabs/Qwen3.8-27B-Pym-IQ2_XXS-GGUF

Qwen3.8-27B-Pym-IQ2_XXS

Precision map

Usage

File / compatibility

SGLabs/
Qwen3.8-27B-Pym-IQ2_XXS-GGUF