Instructions to use xysera/zen-4b with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use xysera/zen-4b with Transformers:

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

pipe = pipeline("text-generation", model="xysera/zen-4b")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("xysera/zen-4b")
model = AutoModelForCausalLM.from_pretrained("xysera/zen-4b", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.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(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use xysera/zen-4b with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "xysera/zen-4b"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "xysera/zen-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/xysera/zen-4b

SGLang

How to use xysera/zen-4b 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 "xysera/zen-4b" \
    --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": "xysera/zen-4b",
		"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 "xysera/zen-4b" \
        --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": "xysera/zen-4b",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use xysera/zen-4b with Docker Model Runner:
```
docker model run hf.co/xysera/zen-4b
```
Browse Quantizations to use this model in llama.cpp, Ollama, LM Studio, or any compatible app.

Zen 4B

Zen 4B is Xysera's public research model for structured reasoning, instruction following, and deliberate problem solving.

This release contains a merged BF16 model produced through Xysera's first full parameter-efficient adaptation pass. It is intended for research, evaluation, and local prototyping.

Model summary

Property	Value
Model family	Zen
Scale	Approximately 4B parameters
Architecture	Decoder-only causal language model
Weight format	BF16 safetensors
Primary language	English
Primary focus	Structured reasoning and instruction following
Release status	Public research release

Training approach

Zen 4B was developed using a proprietary model optimization and adaptation pipeline, incorporating iterative training, checkpoint-based evaluation, and targeted refinement to improve reasoning, instruction following, and structured problem-solving capabilities.

Evaluation

Internal evaluation used 500 examples each from ARC-Easy, HellaSwag, PIQA, and Winogrande. Scores use normalized accuracy where available.

Task	Score
ARC-Easy	68.6%
HellaSwag	61.8%
PIQA	78.8%
Winogrande	73.6%

These are directional internal results, not leaderboard submissions. They should not be compared directly with results using different prompts, shot counts, sample sizes, or evaluation harnesses.

Limitations

The model has not undergone a full safety, bias, or harmful-content evaluation.
Performance is uneven across tasks and may reflect specialization in the adaptation data.
The current evaluation is limited in size.
The model may produce factual errors, incomplete reasoning, or incorrect calculations.
Outputs should be independently verified for consequential use.

Basic usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "xysera/zen-4b"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id, torch_dtype=torch.bfloat16, device_map="auto"
)
messages = [{"role": "user", "content": "Explain how a binary search works."}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.no_grad():
    outputs = model.generate(**inputs, max_new_tokens=256, do_sample=False)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Artifact preservation

The original base checkpoint, QLoRA adapter, intermediate checkpoints, optimizer state, tokenizer, and training logs remain preserved in Xysera's private checkpoint storage. This repository is a clean release repository and does not replace those artifacts.

Status

Zen 4B is an active public research release.

Downloads last month: 747

Safetensors

Model size

4B params

Tensor type

BF16

Model tree for xysera/zen-4b

Quantizations

2 models

xysera
/

zen-4b