Qwen3.8-27B · llama.cpp · RTX 5090 32GB
English | 简体中文 | 日本語 · ← All recipes
Source level: 📣 Claimed · Date: 2026-08-18 · Numbers come from the source. This repo did not measure them again.
Hardware
- RTX 5090 32GB (NVIDIA consumer)
- Reference cost: ≈ $5,400
Model
- Qwen3.8-27B ·
ShapeLearn mixed quantization GPU-1 (2.56 bpw, label IQ2_XXS) - Download: huggingface.co/byteshape/Qwen3.8-27B-GGUF
Engine
- llama.cpp — install: https://github.com/ggml-org/llama.cpp
- Note: llama.cpp (-np 1 -fa on)
Launch command
llama-server -hf byteshape/Qwen3.8-27B-GGUF:<tag> --spec-type draft-mtp --spec-draft-n-max 3
Excerpt from the source (includes ... or placeholders). See the source for full arguments.
Key parameters
- Context: —
- Concurrency: single stream (-np 1)
- Tags:
2bit,dense
Expected speed
- 119.0 tok/s decode. Other tiers on the same card: GPU-2 (2.88 bpw) 111, GPU-3 (3.01 bpw) 108, GPU-4 (3.23 bpw) 104, GPU-5 (3.84 bpw) 94. Qualification: with DFlash2 (7 drafts, text only) GPU-1 176 and GPU-5 167, not the headline (📣 Claimed)
Caveats
- Not a first-hand measurement (claimed or relayed). Your results can differ.
- Speed changes with drivers, engine version, context length, and prompt length.
Source & attribution
- https://huggingface.co/byteshape/Qwen3.8-27B-GGUF
- Credit to the original author. This page only summarizes and links. Data under CC BY 4.0.
Source file: recipes/hf-byteshape-qwen3-8-27b-gguf-1.md