Zen4-Max is a 30B Mixture-of-Experts (MoE) language model with approximately 3B parameters active per forward pass. It delivers high-capability inference at the compute cost of a much smaller dense model, making it ideal for deployments where both quality and throughput matter.

Model Specs

Property	Value
Parameters	30B total / ~3B active (MoE)
Architecture	Sparse MoE transformer (causal LM)
Context window	32,768 tokens
Format	SafeTensors (BF16)
License	Apache 2.0

Zen4 Family

This model is part of the Zen4 model family:

Model	Params	Architecture	Context	Use case
zen4-mini	4B	Dense	40,960	Edge, mobile, low-resource
zen4-pro	14B	Dense	32,768	Professional, complex reasoning
zen4-max	30B MoE	MoE (3B act.)	32,768	High-capability, efficient
zen4-coder	80B MoE	MoE	32,768	Advanced code, 100+ languages
zen4-pro-max	80B MoE	MoE (3B act.)	32,768	Large-scale reasoning

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "zenlm/zen4-max"

tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

messages = [
    {"role": "user", "content": "Design a fault-tolerant microservice architecture for a high-traffic payment system."}
]

text = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer([text], return_tensors="pt").to(model.device)

output_ids = model.generate(**inputs, max_new_tokens=1024)
generated = output_ids[0][len(inputs.input_ids[0]):]
print(tokenizer.decode(generated, skip_special_tokens=True))

Hardware Requirements

MoE models load all expert weights into memory but only activate a subset per token, so total VRAM tracks total parameter count while compute tracks active parameters.

Setup	Minimum VRAM	Recommended
BF16 full precision	64 GB	80 GB (A100/H100)
4-bit quantized	20 GB	24 GB
CPU inference	64 GB RAM	128 GB RAM

Multi-GPU configurations with NVLink or high-bandwidth interconnect are recommended for BF16 full-precision inference.

Formats Available

Format	Location
SafeTensors BF16	`zenlm/zen4-max`
GGUF	`zenlm/zen4-max-gguf` (various quants)

Abliteration

Zen4-Max weights are derived from an abliterated base model. Abliteration removes refusal behavior by identifying and nullifying the "refusal direction" in the model's residual stream.

Method: Directional ablation on the residual stream across all layers Implementation: hanzoai/remove-refusals — Hanzo's production abliteration toolkit Technique: Refusal in LLMs is mediated by a single direction — Arditi et al. Effect: Removes refusal behaviors while preserving all other capabilities Identity layer: Zen identity added via system prompt — full LoRA fine-tuning planned

Abliteration is a feature, not a limitation. It enables unrestricted research, security testing, and applications where safety guardrails are managed at the application layer rather than baked into model weights.

Citation

@misc{zen4max2025,
  title        = {Zen4-Max: 30B Mixture-of-Experts Language Model},
  author       = {{Hanzo AI} and {Zoo Labs Foundation}},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/zenlm/zen4-max}},
  note         = {Apache 2.0 License}
}

Part of the Zen model family by Hanzo AI (Techstars '17) and Zoo Labs Foundation.