Instructions to use defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL with llama.cpp:

Install (macOS, Linux)

curl -LsSf https://llama.app/install.sh | sh
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL
# Run inference directly in the terminal:
llama cli -hf defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL
# Run inference directly in the terminal:
llama cli -hf defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL

Use pre-built binary

# Download pre-built binary from:
# https://github.com/ggerganov/llama.cpp/releases
# Start a local OpenAI-compatible server with a web UI:
./llama-server -hf defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL
# Run inference directly in the terminal:
./llama-cli -hf defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL

Build from source code

git clone https://github.com/ggerganov/llama.cpp.git
cd llama.cpp
cmake -B build
cmake --build build -j --target llama-server llama-cli
# Start a local OpenAI-compatible server with a web UI:
./build/bin/llama-server -hf defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL
# Run inference directly in the terminal:
./build/bin/llama-cli -hf defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL

Use Docker

docker model run hf.co/defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL

LM Studio
Jan

vLLM

How to use defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL

Ollama
How to use defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL with Ollama:
```
ollama run hf.co/defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL
```
Unsloth Desktop

How to use defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL

Configure the model in Pi

# Install Pi:
npm install -g @earendil-works/pi-coding-agent
# Add to ~/.pi/agent/models.json:
{
  "providers": {
    "llama-cpp": {
      "baseUrl": "http://localhost:8080/v1",
      "api": "openai-completions",
      "apiKey": "none",
      "models": [
        {
          "id": "defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL with Docker Model Runner:
```
docker model run hf.co/defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL
```

Lemonade

How to use defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL

Run and chat with the model

lemonade run user.GLM-5.3-DERISKED-UD-Q3_K_XL-UD-Q3_K_XL

List all available models

lemonade list

Hermes Agent

How to use defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL

Configure Hermes

# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL

Configure OpenClaw

# Install OpenClaw:
npm install -g openclaw@latest
# Register the local server and set it as the default model:
openclaw onboard --non-interactive --mode local \
  --auth-choice custom-api-key \
  --custom-base-url http://127.0.0.1:8080/v1 \
  --custom-model-id "defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL:UD-Q3_K_XL" \
  --custom-provider-id llama-cpp \
  --custom-compatibility openai \
  --custom-text-input \
  --accept-risk \
  --skip-health

Run OpenClaw

openclaw agent --local --agent main --message "Hello from Hugging Face"

Access request

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

This repository is public metadata with a manual download gate. Submit an accurate intended-use statement. Approval is discretionary and may be revoked for misuse or redistribution without authorization.

GLM-5.3-DERISKED-UD-Q3_K_XL

Client drop · full GLM-5.3 MoE · Unsloth UD-Q3_K_XL GGUF · Blackfrost DWM a=3

Built by Blackfrost · Las Vegas, Nevada

Public metadata, manual download gate. This is a client delivery of the GLM-5.3 de-risked Q3 GGUF. Request access on the Hub. Do not redistribute the weights without authorization.

What this is

Nine-shard Unsloth UD-Q3_K_XL GGUF of GLM-5.3 (full ~753B MoE, not Flash) after a Blackfrost direction-weight modification (DWM) pass.

The intended behavior is in the weights. It does not depend on a system prompt, LoRA, or a runtime filter. Production DWM details are proprietary and are not disclosed beyond the locked recipe below.

This exact Q3 artifact has not been through the R1-HARMFUL-BENCH-450 judged suite. Do not copy NVFP4/BF16 refusal percentages onto this GGUF.

Specifications


Architecture	`GlmMoeDsaForCausalLM` (GLM-5.3 full, DSA MoE)
Parameters	Full ~753B MoE topology · no expert pruning
Quant	Unsloth `UD-Q3_K_XL` GGUF (mixed K-quants; expert down-proj mostly IQ4_XS / Q6_K)
Artifact size	342,965,977,029 bytes (319.4 GiB)
Shards	9 (`00001-of-00009` … `00009-of-00009`)
Layers	78 main layers + MTP head at blk.78
Hidden size	6,144
Experts	256 routed · top-8 / token · shared expert
Context	Native GLM-5.3 long context (set `--ctx-size` from RAM, not the architectural ceiling)
Runtime	llama.cpp / compatible GGUF loaders
Languages	English and Chinese

Lineage

zai-org/GLM-5.3-BF16
  └─ unsloth/GLM-5.3-GGUF  tree 346b3591c7f28d1a23716f97a065ecf12ec14771
       └─ UD-Q3_K_XL (9 shards, 342,965,977,029 B)
            └─ Blackfrost DWM a=3 skip-early 2  ← this repository


Upstream BF16	`zai-org/GLM-5.3-BF16`
Quant parent	`unsloth/GLM-5.3-GGUF` `UD-Q3_K_XL/`
Quant tree	`346b3591c7f28d1a23716f97a065ecf12ec14771`
Blackfrost change	Weight-level DWM on an independent copy of the Q3 GGUF
Not applied	Extra SFT · DPO · RLHF · expert pruning · GGUF requant

The Unsloth source GGUF was left untouched. This repo is the independent candidate-a3-skip2 writer output.

DWM recipe (locked)


Alpha / scale	3.0
Passes	1
Skip-early	2 (layers 0–1 unchanged)
Norm restore	off (`--no-norm-restore`)
Direction shape	78 × 6,144 f32
Writer	`gguf-dwm-writer` · 8 threads · 2 expert workers
Edited surface	227 packed GGUF tensors · layers 2–77
MTP (blk.78)	not in the 227-target set

Finish markers on the build host: DWM_COMPLETE processed=227 targets=227, SOURCE_UNTOUCHED_CANDIDATE_SIZES_OK, APPLY_OK.

Shard 00001-of-00009 is byte-identical to the Unsloth source (sha256 56d6d59fc554a84c503c2f786e6978d55681e42b8c01370afd923a6286d05e0a) because skip-early leaves the first-shard tensors alone. Later shards match size of the source and are the DWM-edited payload.

Files

file	bytes
`GLM-5.3-UD-Q3_K_XL-00001-of-00009.gguf`	9,428,677
`GLM-5.3-UD-Q3_K_XL-00002-of-00009.gguf`	48,804,973,120
`GLM-5.3-UD-Q3_K_XL-00003-of-00009.gguf`	48,508,432,544
`GLM-5.3-UD-Q3_K_XL-00004-of-00009.gguf`	48,508,432,544
`GLM-5.3-UD-Q3_K_XL-00005-of-00009.gguf`	48,508,432,544
`GLM-5.3-UD-Q3_K_XL-00006-of-00009.gguf`	48,508,432,544
`GLM-5.3-UD-Q3_K_XL-00007-of-00009.gguf`	48,508,432,544
`GLM-5.3-UD-Q3_K_XL-00008-of-00009.gguf`	48,717,290,336
`GLM-5.3-UD-Q3_K_XL-00009-of-00009.gguf`	2,892,122,176
Total	342,965,977,029

Point llama.cpp at shard 00001. It will pull the rest from the same directory.

Load

Recent llama.cpp with GLM-5.3 / GLM-MoE DSA support. Example shape (tune context and GPU layers to the box):

llama-server \
  --model GLM-5.3-UD-Q3_K_XL-00001-of-00009.gguf \
  --ctx-size 32768 \
  --n-gpu-layers 99 \
  --jinja \
  --host 0.0.0.0 --port 8080

Sampling baseline used on the GLM-5.3 line: temperature 1.0, top-p 0.95, thinking on unless you explicitly disable it in the template.

This drop was not load-qualified on Spark after apply. Qualify locally before production: /health, a known-positive completion, and a known-negative control.

What this is not

Not GLM-5.3-Flash (different architecture and GGUF).
Not the BF16 or NVFP4 DERISKED safetensors products.
Not a 450-prompt judged refusal result. Those numbers belong to other checkpoints and must not be cited for this GGUF.
Not a safety-stock model. Output is untrusted. You own access control, logging, and policy.

License and access

Upstream GLM-5.3 remains under Z.AI's GLM-5.3 terms. This derivative is a gated Blackfrost client delivery. Recipients may not republish the weights. Export controls, local law, and the operator's authorization boundary still apply.

Responsible use

Intended for authorized security testing, evaluation, and controlled local inference. Provided as is, without warranty.

Blackfrost · Las Vegas · x.com/Blackfrost_AI

Downloads last month: -

GGUF

Model size

754B params

Architecture

glm-dsa

Hardware compatibility

3-bit

Model tree for defiobi00/GLM-5.3-DERISKED-UD-Q3_K_XL

Base model

zai-org/GLM-5.3-BF16

Quantized

(28)

this model