⚡ Zen LM
Models

zen-max

Trillion-parameter 1.04T MoE open-weights frontier model. Same model as zen4-max.

zen-max

Frontier

The open-weights release of the Zen frontier model. zen-max and zen4-max are the same model -- a 1.04T Mixture-of-Experts architecture with 32B active parameters per token and a 256K context window. zen-max is the brand name; zen4-max is the generation identifier.

Specifications

PropertyValue
Model IDzen-max
Parameters1.04T (32B active)
ArchitectureMoE
Context Window256K tokens
StatusAvailable
HuggingFacezenlm/zen-max

Capabilities

  • Frontier-scale reasoning and analysis
  • 256K context for entire codebases and document collections
  • Complex multi-step problem solving
  • Advanced multilingual understanding (100+ languages)
  • State-of-the-art instruction following
  • Research-grade text generation

Usage

HuggingFace

pip install transformers torch accelerate
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-max")
model = AutoModelForCausalLM.from_pretrained(
    "zenlm/zen-max",
    device_map="auto",
    torch_dtype="auto",
)

inputs = tokenizer("Solve this step by step:", return_tensors="pt").to(model.device)
outputs = model.generate(**inputs, max_new_tokens=1024)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

API

from hanzoai import Hanzo

client = Hanzo(api_key="hk-your-api-key")

response = client.chat.completions.create(
    model="zen-max",
    messages=[{"role": "user", "content": "Derive the proof for the Cauchy-Schwarz inequality."}],
)
print(response.choices[0].message.content)

See Also

  • zen4-max -- Same model, generation identifier
  • zen4-ultra -- zen4 with chain-of-thought reasoning
  • zen4 -- 744B MoE flagship

On this page