emwesoft/GLM-5.3-DFlash2-GGUF overview
GLM 5.3 DFlash2 drafters GGUF GGUF conversions of incoai/GLM 5.3 DFlash2 https://huggingface.co/incoai/GLM 5.3 DFlash2 — the DFlash2 block diffusion drafter fo…
Runs locally from ~2.44 GB disk (4 GB VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
Model Details
Model README
---
license: mit
base_model: incoai/GLM-5.3-DFlash2
library_name: gguf
tags: [gguf, dflash, speculative-decoding, glm, llama.cpp]
---
GLM-5.3 DFlash2 drafters (GGUF)
GGUF conversions of incoai/GLM-5.3-DFlash2 —
the DFlash2 block-diffusion drafter for GLM-5.3 753B
| File | Size |
|---|---|
| GLM-5.3-DFlash2-BF16.gguf | 4.59 GiB — lossless from source |
| GLM-5.3-DFlash2-Q8_0.gguf | 2.44 GiB |
-md GLM-5.3-DFlash2-Q8_0.gguf --spec-type draft-dflash --spec-draft-n-max 7block_size is 8, so 7 is the ceiling — llama.cpp clamps anything higher. Lower values can win
when the target's experts are partly CPU-resident, since each verification step is then more
expensive; sweep it for your setup.
Works with any GLM-5.3 GGUF — DFlash consumes the target's hidden states
(hidden_size 6144, layers [6,20,34,48,62,76]), not its quantisation. Not compatible with
GLM-5.3-Flash (hidden 4096).
Run emwesoft/GLM-5.3-DFlash2-GGUF with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models