zen-eco/instruct/README.md

language, license, tags, base_model, pipeline_tag, library_name

language	license	tags	base_model	pipeline_tag	library_name
en zh	apache-2.0	qwen3 instruct chat assistant zen-ai	Qwen/Qwen3-4B-Instruct	text-generation	transformers

Zen Eco 4B Instruct

Zen Eco Instruct is a 4B parameter instruction-tuned language model optimized for conversational AI and task completion. Based on Qwen3-4B-Instruct, it provides excellent balance between capability and efficiency.

Model Details

• Model Type: Instruction-Tuned Language Model
• Architecture: Qwen3 (4B)
• Parameters: 4 billion
• License: Apache 2.0
• Languages: English, Chinese, and 25+ languages
• Context Length: 32K tokens
• Developed by: Zen AI Team
• Base Model: Qwen/Qwen3-4B-Instruct

Capabilities

• 💬 Conversational: Natural dialogue and chat
• 🎯 Instruction Following: High-quality task completion
• 🌍 Multilingual: 25+ languages supported
• ⚡ Efficient: 28K tokens/sec on RTX 4090
• 📦 Multiple Formats: PyTorch, MLX, GGUF
• 🎛️ Function Calling: Tool use and API integration
• 🔒 Safe: Aligned with safety guidelines

Performance

Throughput

• RTX 4090 (GGUF Q4): 28,000 tokens/sec
• M3 Max (MLX): 32,000 tokens/sec
• RTX 3090 (GGUF Q4): 18,000 tokens/sec

Memory Usage

Format	VRAM/RAM
Q4_K_M	2.5GB
Q8_0	4.2GB
F16	8.0GB

Use Cases

• Conversational AI chatbots
• Task automation and completion
• Content generation and editing
• Code assistance
• Question answering
• Summarization and analysis
• Multilingual applications

Installation

Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "zenlm/zen-eco-4b-instruct",
    torch_dtype="auto",
    device_map="auto"
)
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-eco-4b-instruct")

# Chat
messages = [
    {"role": "system", "content": "You are a helpful assistant."},
    {"role": "user", "content": "Explain quantum computing"}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=500)
print(tokenizer.decode(outputs[0]))

OpenAI API (Zen Engine)

zen-engine serve --model zenlm/zen-eco-4b-instruct --port 3690

from openai import OpenAI

client = OpenAI(
    base_url="http://localhost:3690/v1",
    api_key="not-needed"
)

response = client.chat.completions.create(
    model="zen-eco-4b-instruct",
    messages=[
        {"role": "user", "content": "What is machine learning?"}
    ]
)
print(response.choices[0].message.content)

Training with Zen Gym

Fine-tune for your domain:

cd /path/to/zen-gym

llamafactory-cli train \
    --config configs/zen_eco_instruct_lora.yaml \
    --dataset your_dataset

Benchmarks

Task	Score	Notes
MMLU	58.7%	5-shot
GSM8K	68.2%	8-shot CoT
HumanEval	52.4%	pass@1
MATH	42.8%	4-shot
BBH	61.3%	3-shot

Chat Template

<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
Hello!<|im_end|>
<|im_start|>assistant

Limitations

• May produce incorrect information
• Limited knowledge cutoff
• Quantization reduces quality
• Not suitable for safety-critical applications
• Requires prompt engineering for best results

Citation

@misc{zeneco2025instruct,
  title={Zen Eco 4B Instruct: Efficient Instruction-Tuned Language Model},
  author={Zen AI Team},
  year={2025},
  howpublished={\url{https://huggingface.co/zenlm/zen-eco-4b-instruct}}
}

Links

• GitHub: https://github.com/zenlm
• Zen Gym (Training): https://github.com/zenlm/zen-gym
• Zen Engine (Inference): https://github.com/zenlm/zen-engine

Part of the Zen AI ecosystem.