Zen4 Models

Consumer Line

Dense and MoE models for desktop and edge deployment. All built on abliterated open-source weights.

Zen4 Mini (4B)

Ultra-efficient for edge and mobile. Full quality at 4B parameters.

• HuggingFace | GitHub

Zen4 (8B)

The standard model. Excellent balance of quality and efficiency.

• HuggingFace | GitHub

Zen4 Pro (14B)

Professional-grade for demanding tasks. Strong reasoning and code generation.

• HuggingFace | GitHub

Zen4 Max (30B MoE)

Flagship efficient model. 30B total with only 3B active via MoE.

• HuggingFace | GitHub

Zen4 Pro Max (80B MoE) - Flagship

The ultimate consumer model. Hybrid Gated DeltaNet + Gated Attention + MoE architecture.

• HuggingFace | GitHub

Coder Line

Specialized models for agentic programming and software engineering.

Zen4 Coder Flash (31B MoE)

Fast coding model for rapid iteration. MoE architecture for maximum efficiency.

• HuggingFace | GitHub

Zen4 Coder (80B MoE) - Flagship Code

80B with 512-expert MoE for state-of-the-art code generation and agentic programming.

• HuggingFace | GitHub

Zen4 Coder Pro (355B) - Cloud Only

Dense 355B coding powerhouse for maximum code intelligence.

• HuggingFace | GitHub

Ultra Line

Trillion-parameter models for cloud deployment.

Zen4 Ultra (1.04T MoE) - Cloud

Trillion-parameter frontier model with 384 experts and native vision capabilities.

• HuggingFace | GitHub

Zen4 Ultra Max (1T+ MoE) - Coming Soon

Next-generation trillion-parameter model with 1M context window.

Zen5 — Next Generation

Key Features

• 2T+ parameters — the largest open-weight model ever released
• MoDE architecture — Mixture of Distilled Experts for efficient routing
• On-chain training — verifiable training via NVIDIA TEE on hanzo.network
• 1M+ context — full codebase and document understanding
• GT-QLoRA — Gate-Targeted fine-tuning for MoE behavioral modification (paper)

Request Research Access

Zen5 is in private research preview. Researchers and institutions can request early access to preprints, weights, and evaluation under a special research license.

• Request Research Access — for researchers and institutions
• Join Waitlist — for API access when available
• GT-QLoRA Paper — our MoE fine-tuning research

Formats

All locally-runnable models are available in multiple formats:

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer

# Load any Zen4 model
model = AutoModelForCausalLM.from_pretrained("zenlm/zen4-pro-max")
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen4-pro-max")

messages = [{"role": "user", "content": "Hello!"}]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Multimodal & Specialized

In addition to Zen4, the Zen LM family includes multimodal and specialized models:

All models available at huggingface.co/zenlm.