zen-agent-4b

4B parameter tool-calling agent with Model Context Protocol (MCP) support

Model Details

Developed by: Zen Research Authors
Organization: Zen Research DAO under Zoo Labs Inc (501(c)(3) Non-Profit)
Location: San Francisco, California, USA
Model type: language-model
Architecture: Qwen3-4B with MCP
Parameters: 4B
License: Apache 2.0
Training: Trained with Zen Gym
Inference: Optimized for Zen Engine

🌟 Zen AI Ecosystem

This model is part of the Zen Research hypermodal AI family - the world's most comprehensive open-source AI ecosystem.

Complete Model Family

Language Models:

zen-nano-0.6b - 0.6B edge model (44K tokens/sec)
zen-eco-4b-instruct - 4B instruction model
zen-eco-4b-thinking - 4B reasoning model
zen-agent-4b - 4B tool-calling agent

3D & World Generation:

zen-3d - Controllable 3D asset generation
zen-voyager - Camera-controlled world exploration
zen-world - Large-scale world simulation

Video Generation:

zen-director - Text/image-to-video (5B)
zen-video - Professional video synthesis
zen-video-i2v - Image-to-video animation

Audio Generation:

zen-musician - Music generation (7B)
zen-foley - Video-to-audio Foley effects

Infrastructure:

Zen Gym - Unified training platform
Zen Engine - High-performance inference

Usage

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("zenlm/zen-agent-4b")
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-agent-4b")

inputs = tokenizer("Hello!", return_tensors="pt")\noutputs = model.generate(**inputs)\nprint(tokenizer.decode(outputs[0]))

With Zen Engine

# High-performance inference (44K tokens/sec on M3 Max)
zen-engine serve --model zenlm/zen-agent-4b --port 3690

# OpenAI-compatible API
from openai import OpenAI

client = OpenAI(base_url="http://localhost:3690/v1")
response = client.chat.completions.create(
    model="zenlm/zen-agent-4b",
    messages=[{"role": "user", "content": "Hello!"}]
)

Training

Fine-tune with Zen Gym:

git clone https://github.com/zenlm/zen-gym
cd zen-gym

# LoRA fine-tuning
llamafactory-cli train --config configs/zen_lora.yaml \
    --model_name_or_path zenlm/zen-agent-4b

# GRPO reinforcement learning (40-60% memory reduction)
llamafactory-cli train --config configs/zen_grpo.yaml \
    --model_name_or_path zenlm/zen-agent-4b

Supported methods: LoRA, QLoRA, DoRA, GRPO, GSPO, DPO, PPO, KTO, ORPO, SimPO, Unsloth

Performance

Speed: 28K tokens/sec (RTX 4090)
Memory: 2.5GB (Q4_K_M) to 8GB (F16)
MCP: Full Model Context Protocol support
Tools: 100+ function calling accuracy

Ethical Considerations

Open Research: Released under Apache 2.0 for maximum accessibility
Environmental Impact: Optimized for eco-friendly deployment
Transparency: Full training details and model architecture disclosed
Safety: Comprehensive testing and evaluation
Non-Profit: Developed by Zoo Labs Inc (501(c)(3)) for public benefit

Citation

@misc{zenzenagent4b2025,
  title={zen-agent-4b: 4B parameter tool-calling agent with Model Context Protocol (MCP) support},
  author={Zen Research Authors},
  year={2025},
  publisher={Zoo Labs Inc},
  organization={Zen Research DAO},
  url={https://huggingface.co/zenlm/zen-agent-4b}
}

License

Apache License 2.0

Zen Research - Building open, eco-friendly AI for everyone 🌱

Downloads last month: 215

Safetensors

Model size

4.02B params

Tensor type

F16

zenlm
/

zen-agent-4b