Zen Agent is a 4B parameter model specialized for agentic workflows, tool use, and function calling. Built on Qwen3-4B and fine-tuned for autonomous task execution with Model Context Protocol (MCP) support.

Training Data

This model was fine-tuned on:

Salesforce/xlam-function-calling-60k - 60K high-quality function calling examples

Tool Calling Capabilities

Function detection and parameter extraction
Multi-tool coordination
JSON schema understanding
Structured output generation
Chain-of-thought tool reasoning

Available Formats

PyTorch (Default)

model.safetensors.index.json - Model index
model-00001-of-00002.safetensors - Model shard 1
model-00002-of-00002.safetensors - Model shard 2

MLX Format (Apple Silicon)

mlx/ - Quantized MLX format for Apple Silicon (Qwen3-4B)
Optimized for M1/M2/M3 Macs
4.5 bits per weight quantization

GGUF Format (llama.cpp)

gguf/zen-eco-4b-agent-qwen3.gguf - Q4_K_M (2.3GB)
gguf/zen-eco-4b-agent-Q2_K.gguf - Q2_K (1.6GB)
gguf/zen-eco-4b-agent-f16.gguf - F16 (7.5GB)

Usage

Tool Calling Example

from transformers import AutoModelForCausalLM, AutoTokenizer

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

prompt = """
Available functions:
- get_weather(location: str) -> dict
- search_web(query: str) -> list

User: What's the weather in Paris?
Assistant:"""

inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
print(tokenizer.decode(outputs[0]))

MLX (Apple Silicon)

from mlx_lm import load, generate

model, tokenizer = load("zenlm/zen-eco-4b-agent")
response = generate(model, tokenizer, prompt="Call weather API for Tokyo", max_tokens=50)

GGUF (LM Studio / llama.cpp)

# Download
wget https://huggingface.co/zenlm/zen-eco-4b-agent/resolve/main/gguf/zen-eco-4b-agent-qwen3.gguf

# Run with llama.cpp
./llama-cli -m zen-eco-4b-agent-qwen3.gguf -p "Get weather for Paris"

Model Details

Base: Qwen3-4B (NOT Qwen2)
Parameters: 4B
Context Length: 32K tokens
Training: Fine-tuned on xlam-function-calling-60k dataset
License: Apache 2.0

Performance

Throughput

RTX 4090 (GGUF Q4): 28,000 tokens/sec
M3 Max (MLX): 30,000 tokens/sec
RTX 3090 (GGUF Q4): 18,000 tokens/sec

Benchmarks

Task	Score	Notes
ToolBench	72.4%	Tool selection accuracy
APIBench	68.9%	API calling correctness
MMLU	56.3%	5-shot general knowledge

Citation

@misc{zen-agent-4b-2025,
  title={Zen Agent 4B: Tool-Calling Language Model},
  author={Zen AI Team},
  year={2025},
  howpublished={\url{https://huggingface.co/zenlm/zen-agent-4b}}
}

@article{xlam2024,
  title={xLAM: A Family of Large Action Models to Empower AI Agent Systems},
  author={Salesforce Research},
  journal={arXiv preprint},
  year={2024}
}

Part of the Zen AI ecosystem.

README.md

Zen Agent 4B