Zen4-Pro-Max is an 80B Mixture-of-Experts (MoE) language model with approximately 3B parameters active per forward pass. It is the flagship model of the Zen4 family, designed for large-scale reasoning, complex multi-step tasks, and demanding professional workloads where maximum capability is required.

This model is part of the Zen4 model family:

from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "zenlm/zen4-pro-max"

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

messages = [
    {
        "role": "user",
        "content": (
            "Derive the time complexity of Dijkstra's algorithm using a Fibonacci heap, "
            "then compare it with a binary heap implementation across sparse and dense graphs."
        ),
    }
]

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=2048)
generated = output_ids[0][len(inputs.input_ids[0]):]
print(tokenizer.decode(generated, skip_special_tokens=True))

from mlx_lm import load, generate

model, tokenizer = load("zenlm/zen4-pro-max-mlx")

response = generate(
    model,
    tokenizer,
    prompt="Explain the formal semantics of linear type systems.",
    max_tokens=1024,
    verbose=True,
)

The MoE architecture activates only ~3B parameters per token, meaning compute cost scales with active parameters while memory scales with total parameters.

Zen4-Pro-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.

@misc{zen4promax2025,
  title        = {Zen4-Pro-Max: 80B MoE Flagship Language Model for Large-Scale Reasoning},
  author       = {{Hanzo AI} and {Zoo Labs Foundation}},
  year         = {2025},
  howpublished = {\url{https://huggingface.co/zenlm/zen4-pro-max}},
  note         = {Apache 2.0 License}
}

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