emwesoft/GLM-5.3-NVFP4-GGUF overview
GLM 5.3 753B — NVFP4 GGUF no MTP Native GGUF of incoai/GLM 5.3 NVFP4 https://huggingface.co/incoai/GLM 5.3 NVFP4 , the vendor's ModelOpt NVFP4 repack of zai or…
Runs locally from ~34.88 GB disk (32 GB+ VRAM class GPUs with llama.cpp / guIDE).
Repository Files & Downloads
| File | Type | Quantization | Size | Link |
|---|---|---|---|---|
| GLM-5.3-NVFP4-00001-of-00010.gguf | GGUF | GGUF | 42.20 GB | Download |
| GLM-5.3-NVFP4-00002-of-00010.gguf | GGUF | GGUF | 42.19 GB | Download |
| GLM-5.3-NVFP4-00003-of-00010.gguf | GGUF | GGUF | 42.19 GB | Download |
| GLM-5.3-NVFP4-00004-of-00010.gguf | GGUF | GGUF | 42.19 GB | Download |
| GLM-5.3-NVFP4-00005-of-00010.gguf | GGUF | GGUF | 42.19 GB | Download |
| GLM-5.3-NVFP4-00006-of-00010.gguf | GGUF | GGUF | 42.19 GB | Download |
| GLM-5.3-NVFP4-00007-of-00010.gguf | GGUF | GGUF | 42.19 GB | Download |
| GLM-5.3-NVFP4-00008-of-00010.gguf | GGUF | GGUF | 42.19 GB | Download |
| GLM-5.3-NVFP4-00009-of-00010.gguf | GGUF | GGUF | 42.19 GB | Download |
| GLM-5.3-NVFP4-00010-of-00010.gguf | GGUF | GGUF | 34.88 GB | Download |
Model Details
| Model ID | emwesoft/GLM-5.3-NVFP4-GGUF |
|---|---|
| Author | emwesoft |
| Pipeline | text-generation |
| License | mit |
| Base model | zai-org/GLM-5.3,incoai/GLM-5.3-NVFP4 |
| Last modified | 2026-08-29T10:34:33.000Z |
Model README
---
license: mit
base_model: [zai-org/GLM-5.3, incoai/GLM-5.3-NVFP4]
pipeline_tag: text-generation
library_name: gguf
tags: [gguf, nvfp4, glm, glm-dsa, llama.cpp]
---
GLM-5.3 753B — NVFP4 GGUF (no MTP)
Native GGUF of incoai/GLM-5.3-NVFP4, the
vendor's ModelOpt NVFP4 repack of zai-org/GLM-5.3
(753B, glm-dsa, 78 blocks, hidden 6144, 256 experts). The NVFP4 weights are kept as NVFP4 —
this is not a requantisation.
445 GB, 10 shards. block_count 78, 1947 tensors, 225 NVFP4 tensors.
Related repos
| | |
|---|---|
| With MTP head | emwesoft/GLM-5.3-NVFP4-MTP-GGUF — adds block 78, enables --spec-type draft-mtp |
| DFlash2 drafters | emwesoft/GLM-5.3-DFlash2-GGUF — speculative decoding for this model |
This variant has no MTP head, so --spec-type draft-mtp is unavailable. Use the DFlash2
drafters for speculative decoding, or the MTP repo above.
Engine requirements
llama.cpp with glm-dsa + GGML_TYPE_NVFP4. Three fixes are not yet upstream:
- jinja numeric attribute access (
obj.0) — GLM-5.3's chat template uses
m.content.0.output. Without it the template throws, caps_get() swallows it,
supports_tool_calls reports false, and every tool call comes back as plain text.
glm-dsalayer-input exposure — DFlash needsres->t_layer_inp[il]; without it
attaching a drafter aborts on the first decode with GGML_ASSERT(t_layer_inp[il] != nullptr).
- Whitespace tolerance before
</tool_call>— a stray newline makes the streaming
parser recognise a tool call then lose it, aborting from compute_diffs.
Measured throughput
2x RTX PRO 6000 Blackwell + 4x RTX 3090 + 251 GB RAM, 400K context, -t 36 -tb 40,
experts partly CPU-resident (the weights do not fit in 288 GB of VRAM):
| config | acceptance | decode |
|---|---|---|
| MTP head, n-max 3 | 74.9% (mean len 3.24) | 9.3-12.2 tok/s |
| DFlash2 Q8_0, n-max 4 | 67.7% (mean len 3.69) | 7.6-12.8 tok/s |
| no speculation | - | ~10 tok/s |
Throughput is prompt-dependent because acceptance is. Threads matter: on a 24-core/48-thread
CPU, -t 48 collapsed decode to 0.5 tok/s — the ggml threadpool busy-spins and starves the CUDA
submission thread. Leave headroom.
Sampling
From generation_config.json: temperature 1.0, top_p 0.95. The template exposes
low/high/max reasoning effort only; anything else becomes max, and thinking cannot be
disabled.
Run emwesoft/GLM-5.3-NVFP4-GGUF with guIDE
Download guIDE — the AI-native code editor with local LLM inference and 69 built-in tools.
Source: Hugging Face · Compare models