zenlm/zen-eco-instruct

Zen Eco Instruct

Model Description

Balanced 4B model for consumer hardware

Base Model: Qwen/zen-3B-Instruct
Parameters: 4B
Architecture: zenForCausalLM
Context Length: 32,768 tokens
Training Framework: Zoo-Gym v2.0.0 with RAIS

🎉 v1.0.1 Release (2025)

Recursive Self-Improvement Update

This release introduces our groundbreaking Recursive AI Self-Improvement System (RAIS), where models learn from their own work sessions.

Key Metrics:

• 📊 94% effectiveness across 20 training examples
• 🔒 Enhanced security and error handling
• 📚 Improved documentation understanding
• 🎯 Stronger model identity

What's New

• Security: Fixed API token exposure, added path validation
• Documentation: Hierarchical structure, comprehensive guides
• Identity: Clear branding, no base model confusion
• Technical: Multi-format support (MLX, GGUF, SafeTensors)
• Learning: Pattern recognition from real work sessions

Installation

Using Transformers

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("zenlm/zen-eco-instruct")
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-eco-instruct")

# Generate text
inputs = tokenizer("Hello, how are you?", return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=100)
response = tokenizer.decode(outputs[0], skip_special_tokens=True)
print(response)

Using MLX (Apple Silicon)

from mlx_lm import load, generate

model, tokenizer = load("zenlm/zen-eco-instruct")
response = generate(model, tokenizer, "Hello, how are you?", max_tokens=100)
print(response)

Using llama.cpp

# Download GGUF file
wget https://huggingface.co/zenlm/zen-eco-instruct/resolve/main/zen-eco-instruct-Q4_K_M.gguf

# Run inference
./llama.cpp/main -m zen-eco-instruct-Q4_K_M.gguf -p "Hello, how are you?" -n 100

Training with Zoo-Gym

This model supports fine-tuning with zoo-gym:

from zoo_gym import ZooGym

gym = ZooGym("zenlm/zen-eco-instruct")
gym.train(
    dataset="your_data.jsonl",
    epochs=3,
    use_lora=True,
    lora_r=32,
    lora_alpha=64
)

# Enable recursive improvement
gym.enable_recursive_improvement(
    feedback_threshold=0.85,
    improvement_cycles=5
)

Model Formats

This model is available in multiple formats:

• SafeTensors: Primary format for transformers
• GGUF: Quantized formats (Q4_K_M, Q5_K_M, Q8_0)
• MLX: Optimized for Apple Silicon (4-bit, 8-bit)
• ONNX: For edge deployment

Performance

Benchmark	Score
MMLU	51.7%
GSM8K	32.4%
HumanEval	22.6%
HellaSwag	76.4%

Inference Speed:

• Apple M2 Pro: 45-52 tokens/second
• RTX 4090: 120-140 tokens/second
• CPU (i7-12700K): 8-12 tokens/second

Environmental Impact

• Energy Usage: 95% less than 70B models
• CO₂ Saved: ~1kg per user per month
• Memory: 4.0GB (FP16)

Citation

@misc{zen_v1_0_1_2025,
    title={zen-eco-instruct: Efficient Language Model for Edge Deployment},
    author={Hanzo AI and Zoo Labs Foundation},
    year={2025},
    version={1.0.1}
}

Partnership

Built by Hanzo AI (Techstars-backed) and Zoo Labs Foundation (501(c)(3) non-profit) for open, private, and sustainable AI.

---

© 2025 • Built with ❤️ by Hanzo AI & Zoo Labs Foundation