Zen 5 — Next Generation
Fifth-generation agentic models with MoDE (Mixture of Distilled Experts) and native chain-of-thought reasoning
zen5
TBAAgentic frontier model with native chain-of-thought
zen5-pro
TBAHigh-throughput agentic model for production
zen5-max
TBAMaximum context for document-scale analysis
zen5-ultra
TBADeepest reasoning with self-verification
zen5-mini
TBAEfficient agentic model, zen5-class intelligence
Zen 4 — Flagship
Fourth-generation flagship models with MoE architecture, long context, and state-of-the-art benchmarks
zen4
744B MoE (40B active)Flagship model for complex reasoning
zen4-ultra
744B MoE (40B active)Maximum reasoning with chain-of-thought
zen4-pro
80B MoE (3B active)High-capability efficient MoE
zen4-max
DenseMost capable model, agentic tasks
zen4.6
DenseExtended context, cost efficient
zen4-mini
DenseUltra-fast, free tier
zen4-thinking
80B MoE (3B active)Dedicated reasoning
Zen 4 Coder
Specialized code models for software engineering, generation, review, and debugging
zen4-coder
480B MoE (35B active)Code generation, review, debugging
zen4-coder-pro
480B Dense BF16Full-precision code analysis
zen4-coder-flash
30B MoE (3B active)Fast inline completions
Zen 3 — Multimodal
Third-generation multimodal models spanning text, vision, audio, and safety
zen3-omni
~200B Dense MultimodalText, vision, audio
zen3-vl
30B MoE (3B active)Vision-language
zen3-nano
8B DenseEdge deployment, free tier
zen3-guard
4B DenseContent safety, 119 languages
Zen 3 Image
Diffusion models for image generation, editing, and creative workflows
zen3-image
DiffusionBest general-purpose image generation
zen3-image-max
DiffusionMaximum quality, professional creative
zen3-image-dev
DiffusionDevelopment and iteration
zen3-image-fast
DiffusionReal-time generation
zen3-image-sdxl
Diffusion1024px high-resolution
zen3-image-playground
DiffusionAesthetic artistic generation
zen3-image-ssd
1B DiffusionFastest diffusion model
zen3-image-jp
DiffusionJapanese-specialized
Zen 3 Audio & Speech
Transcription, text-to-speech, real-time ASR, and broadcast-quality audio
zen3-audio
1.5B ASR100+ language transcription
zen3-audio-fast
809M ASRFastest transcription
zen3-asr
Streaming ASRReal-time <500ms latency
zen3-asr-v1
Streaming ASRLegacy compatibility
zen3-tts
82M TTS40+ voices, 8 languages
zen3-tts-hd
TTS HDBroadcast-quality 48kHz
zen3-tts-fast
82M TTSLow-latency voice agents
Zen 3 Embedding
Dense embeddings for semantic search, RAG, classification, and retrieval
zen3-embedding
3072 dimensionsRAG, search, classification
zen3-embedding-medium
4BCost-effective retrieval
zen3-embedding-small
0.6BHigh-throughput
zen3-embedding-openai
3072 dimensionsOpenAI-compatible drop-in
Zen 3 Reranker
Cross-encoder rerankers for improving search and RAG pipeline accuracy
Capabilities Matrix
Each model specializes in different modalities and tasks
| Model | Text | Image | Video | Audio | 3D | Code | Agents |
|---|---|---|---|---|---|---|---|
| zen5 | ✓ | — | — | — | — | ✓ | ✓ |
| zen4 | ✓ | — | — | — | — | ✓ | ✓ |
| zen4-max | ✓ | — | — | — | — | ✓ | ✓ |
| zen4-ultra | ✓ | — | — | — | — | ✓ | ✓ |
| zen4-coder | ✓ | — | — | — | — | ✓ | ✓ |
| zen3-omni | ✓ | ✓ | — | ✓ | — | — | — |
| zen3-vl | ✓ | ✓ | — | — | — | — | — |
| zen3-nano | ✓ | — | — | — | — | ✓ | — |
| zen3-guard | ✓ | — | — | — | — | — | — |
| zen3-image | ✓ | ✓ | — | — | — | — | — |
| zen3-audio | ✓ | — | — | ✓ | — | — | — |
| zen3-tts | ✓ | — | — | ✓ | — | — | — |
| zen3-embedding | ✓ | — | — | — | — | — | — |
| zen3-reranker | ✓ | — | — | — | — | — | — |
Infrastructure
Production-ready tools for training and deploying Zen models
Hanzo Engine
High-performance cloud inference — 60+ architectures, CUDA/Metal, OpenAI-compatible API
Hanzo Edge
On-device AI inference — run models locally on any device, browser, or embedded system
Zen Gym
Unified training platform for all Zen models with LoRA, QLoRA, GRPO, and more
Hanzo MCP
Model Context Protocol for AI context management and tool use
Quick Start
Get started with any Zen model in seconds
# Install and run any model
pip install transformers torch
# Use directly
from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("zenlm/zen-eco-4b-instruct")
tokenizer = AutoTokenizer.from_pretrained("zenlm/zen-eco-4b-instruct")
# Or use via Hanzo Cloud API
from hanzoai import Hanzo
client = Hanzo(api_key="hk-your-api-key")
response = client.chat.completions.create(
model="zen-eco-4b-instruct",
messages=[{"role": "user", "content": "Hello!"}]
)