Instructions to use aj9o9/GLM-5.3-Flash-GGUF with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Notebooks
Google Colab
Kaggle
Local Apps Settings

How to use aj9o9/GLM-5.3-Flash-GGUF 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 aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS
# Run inference directly in the terminal:
llama cli -hf aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS

Install from WinGet (Windows)

winget install llama.cpp
# Start a local OpenAI-compatible server with a web UI:
llama serve -hf aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS
# Run inference directly in the terminal:
llama cli -hf aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS

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 aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS
# Run inference directly in the terminal:
./llama-cli -hf aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS

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 aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS
# Run inference directly in the terminal:
./build/bin/llama-cli -hf aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS

Use Docker

docker model run hf.co/aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS

LM Studio
Jan

vLLM

How to use aj9o9/GLM-5.3-Flash-GGUF with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "aj9o9/GLM-5.3-Flash-GGUF"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "aj9o9/GLM-5.3-Flash-GGUF",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS

Ollama
How to use aj9o9/GLM-5.3-Flash-GGUF with Ollama:
```
ollama run hf.co/aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS
```
Unsloth Desktop

How to use aj9o9/GLM-5.3-Flash-GGUF with Pi:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS

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": "aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS"
        }
      ]
    }
  }
}

Run Pi

# Start Pi in your project directory:
pi

Docker Model Runner
How to use aj9o9/GLM-5.3-Flash-GGUF with Docker Model Runner:
```
docker model run hf.co/aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS
```

Lemonade

How to use aj9o9/GLM-5.3-Flash-GGUF with Lemonade:

Pull the model

# Download Lemonade from https://lemonade-server.ai/
lemonade pull aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS

Run and chat with the model

lemonade run user.GLM-5.3-Flash-GGUF-IQ2_XXS

List all available models

lemonade list

Hermes Agent

How to use aj9o9/GLM-5.3-Flash-GGUF with Hermes Agent:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS

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 aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS

Run Hermes

hermes

Atomic Chat

OpenClaw

How to use aj9o9/GLM-5.3-Flash-GGUF with OpenClaw:

Start the llama.cpp server

# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS

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 "aj9o9/GLM-5.3-Flash-GGUF:IQ2_XXS" \
  --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"

GLM-5.3-Flash AJ-IQ2_XXS, 87.3G, 2.18 bpw

Hand-mixed --tensor-type-file bake of GLM-5.3-Flash (320.6B total / 17.3B active). Beats Unsloth UD-IQ1_S by 26% PPL at 5.7G smaller. Still costs 2.49× BF16.

Runs on a 24G card with 64G system RAM via -cmoe. Measured, see below. Needs llama.cpp at 1f817ef or later from PR #27752, unmerged. Earlier commits return nan on the CPU expert path.

PPL vs BF16

Corpus: Wikitext-2 wiki.test.raw, 8 chunks, n_ctx=512. Tool: llama-perplexity -ngl 99 -t 64, all weights on GPU. PR #27752 @ 20de49688. One build, one corpus, matched chunk count. Only quantization differs.

model	size	PPL	vs BF16
BF16	642G	5.0739 ± 0.307	base
AJ-IQ2_XXS	87.3G	12.6296 ± 0.845	2.49×
Unsloth UD-IQ1_S	93.1G	17.0427 ± 1.176	3.36×

Per chunk:

chunk	1	2	3	4	5	6	7	8
BF16	1.9119	7.2195	6.2120	4.8322	3.7121	3.9721	4.1087	5.0739
AJ-IQ2_XXS	2.9418	10.9033	15.1379	12.5884	9.6512	10.3919	10.9135	12.6296
UD-IQ1_S	3.3618	14.4338	17.4792	15.6171	12.2137	13.3059	13.7524	17.0427

AJ-IQ2_XXS 565-chunk full run: 17.2454 ± 0.14367.

Consumer hardware: needs `1f817ef` or later

Runs on RTX 3090 (24G) + 64G system RAM with experts on CPU. Verified end to end:

llama-perplexity -m GLM-5.3-Flash-AJ-IQ2_XXS-00001-of-00002.gguf \
  -f wiki.test.raw -ngl 99 -cmoe -t 12 --chunks 8
[1]3.0012,[2]11.4442,[3]16.9122,[4]12.9369,[5]9.8740,[6]10.6446,[7]10.8932,[8]13.0209,
Final estimate: PPL = 13.0209 +/- 0.87912

13.0209 on 3090 CPU-offload vs 12.6296 on a 96G card all-GPU. 3.1% apart, error bars overlap. The two backends agree. 2:45 load, 434 s/pass, 8 chunks in ~7 min.

Commits before 1f817ef return nan from chunk 2 on this path. Reproduced on AJ-IQ2_XXS and on unsloth/GLM-5.3-Flash-UD-IQ1_S; BF16 through the same path was clean, so it hit quantized MoE tensors specifically. Chunk 1 matched the GPU path to 4 digits (3.0456 vs 3.0468), then nan. Fixed upstream by 1f817ef "skip foreign-sequence cells when building the glm5next k-pools". Check your build.

Speed

llama-bench, same 87.3G file on every row. Where the routed experts live is the only variable that matters.

Single 3090 + 64GB DDR5 RAM

/build/bin/llama-bench   --model GLM-5.3-Flash-IQ2XXS-00001-of-00002.gguf   --n-gpu-layers 99   --n-cpu-moe 40   -fa on  -t 12  -ctk q8_0 -ctv q8_0  -p 2048 -n 128 -r 2

model	size	params	backend	ngl	n_cpu_moe	type_k	type_v	fa	test	t/s
glm5next 312B.A17B IQ2_XXS - 2.0625 bpw	81.34 GiB	313.33 B	CUDA	99	40	q8_0	q8_0	1	pp2048	36.69 ± 6.23
glm5next 312B.A17B IQ2_XXS - 2.0625 bpw	81.34 GiB	313.33 B	CUDA	99	40	q8_0	q8_0	1	tg128	11.55 ± 0.68

Run

90G+ VRAM, everything on GPU:

llama-cli -m GLM-5.3-Flash-AJ-IQ2_XXS-00001-of-00002.gguf -ngl 99 -c 32768

24G card + 64G system RAM, experts on CPU:

llama-cli -m GLM-5.3-Flash-AJ-IQ2_XXS-00001-of-00002.gguf -ngl 99 -cmoe -c 32768

Point at shard 1, llama.cpp finds shard 2. -ncmoe N offloads only N layers if you have VRAM to spare.

Context is cheap. MLA caches a 512-wide latent over 11 layers, the 34 KDA layers hold constant-size state: 32K = 0.44G, 131K = 1.55G, 300K = 3.45G.

Files

model: AJ-IQ2_XXS/GLM-5.3-Flash-AJ-IQ2_XXS-0000{1,2}-of-00002.gguf
mix: glm53-flash-iq2xxs.txt
base: zai-org/GLM-5.3-Flash, MIT
BF16 GGUF + imatrix: unsloth/GLM-5.3-Flash-GGUF. Saved a 328G FP8 download and an imatrix pass that is not practical at this scale.

Downloads last month: -

GGUF

Model size

313B params

Architecture

glm5next

Hardware compatibility

2-bit

Model tree for aj9o9/GLM-5.3-Flash-GGUF

Base model

zai-org/GLM-5.3-Flash

Quantized

(66)

this model

GLM-5.3-Flash AJ-IQ2_XXS, 87.3G, 2.18 bpw

PPL vs BF16

Consumer hardware: needs 1f817ef or later

Speed

Run

Files

Model tree for aj9o9/GLM-5.3-Flash-GGUF

Consumer hardware: needs `1f817ef` or later