Instructions to use AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8 with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
pipe(text=messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoProcessor, AutoModelForMultimodalLM

processor = AutoProcessor.from_pretrained("AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8")
model = AutoModelForMultimodalLM.from_pretrained("AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8", device_map="auto")
messages = [
    {
        "role": "user",
        "content": [
            {"type": "image", "url": "https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/p-blog/candy.JPG"},
            {"type": "text", "text": "What animal is on the candy?"}
        ]
    },
]
inputs = processor.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(processor.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8

SGLang

How to use AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8 with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8 with Docker Model Runner:
```
docker model run hf.co/AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

GLM-5.3-Flash-AZ-Focus

GLM-5.3-Flash with the second-guessing turned down: 40-50% less thinking at max effort, with reasoning, maths and code accuracy level with or above the original.

GLM-5.3-Flash thinks carefully, but a lot of that thinking is hesitation: "Wait…", "Let me double-check…", re-deriving answers it already had. We optimised the model's weights against that habit, while leaving alone the checking it does as an agent (running tests, inspecting output before it submits). AZ-Focus still reasons step by step. It just loops less.

About half the thinking at max effort on knowledge and reasoning questions, at the same accuracy (FP8, measured on this build).
Agentic coding close to the original: SWE-bench Verified 94 / 93 vs 96 (first 100 tasks).
Maths and science level or up on the EXL3 build of the same edit (AIME 2025, GPQA Diamond).
Weights only. The edit lives in the weights; no adapter, runtime patch or extra code.

This repository is FP8, in the same format and file layout as Z.AI's official FP8 release: a drop-in replacement for serving with vLLM or SGLang. It uses a slightly lighter setting of the edit than the 4-bit EXL3 version at AxonZeta/GLM-5.3-Flash-AZ-Focus-EXL3, tuned so agentic coding stays close to the original at FP8 precision.

Made by AxonZeta, an independent AI lab in Manchester, UK.

Results (max effort)

This FP8 build, against Z.AI's FP8 release, both served with vLLM on the same hardware:

Benchmark	GLM-5.3-Flash FP8	AZ-Focus FP8	Thinking tokens
Reasoning and knowledge (48 questions: 24 MMLU-Pro, 24 BIG-Bench Hard)	35/48 · 2,351 tokens	35/48 · 1,169 tokens	50% less
SWE-bench Verified, first 100 (mini-SWE-agent)	96	94 / 93 (two runs)

The same edit on the 4-bit EXL3 build (a slightly stronger setting), for a wider picture:

Benchmark	GLM-5.3-Flash	AZ-Focus EXL3	Thinking tokens
MMLU-Pro (500 questions)	86.2% · 2,114 tokens	86.8% · 1,289 tokens	39% less
BIG-Bench Hard (540 tasks)	83.7% · 634	84.1% · 326	49% less
AIME 2025 (30, 2 runs)†	75.0% · 16.5k	81.7% · 11.2k	32% less
GPQA Diamond (198)†	80.8%	82.3%	~20% less
HumanEval+ & MBPP+ (EvalPlus)†	85.8%	86.9%	~40% less
τ²-bench airline / retail (two runs)†	0.89 / 0.919	0.88 / 0.905

Thinking on, Z.AI's recommended sampling, same serving setup for both models in each table. Single runs move by about a point either way (a few tasks on SWE-bench). SWE-bench runs used a 12-hour limit per task sandbox. None of these benchmarks was used to build the model. † Measured on an earlier build of the same edit.

Effort setting

AZ-Focus works best at max or high effort, where most of the hesitation is. At low effort GLM-5.3-Flash already thinks very little, and the edit costs accuracy without saving much (BBH 80.0% vs 83.5%, MMLU-Pro 83.0% vs 84.2%). Use the original if you run at low effort.

Usage

Use it exactly like Z.AI's GLM-5.3-Flash FP8 release: same architecture, tokenizer, chat template, file layout and size, served with the same engines and settings. No adapter, patch or extra code. The difference shows in thinking mode; with thinking off, output length is unchanged.

Limitations

Agentic coding is about 2-3 points below the original on SWE-bench Verified (two runs of 100 tasks). If long agent runs are your main use, compare both on your own tasks.
Low effort: see above.
Other languages, very long contexts and multi-hour agent sessions haven't been measured separately yet.

License

Questions, or want to build on this with us? hello@axonzeta.ai

Downloads last month: 144

Safetensors

Model size

321B params

Tensor type

BF16

F8_E4M3

F32

Model tree for AxonZeta/GLM-5.3-Flash-AZ-Focus-FP8

Base model

zai-org/GLM-5.3-Flash

Finetuned

(35)

this model

Quantizations

1 model