# Hanzo AI - AI Engineering Platform > Build, deploy, and scale AI applications with Hanzo AI engineering platform. # The AI cloudyou build on. Every model behind one key. Agents that call real tools. The search, storage and serving underneath. Type below to open a chat with it. Ask Hanzo Write Learn Code Preview ### Some work needs your actual files Hanzo Dev opens the repo you are in, makes the change, and runs the tests. It is in the macOS app. [Download](https://hanzo.app/download)[Learn more](https://hanzo.ai/dev) ## What is in the cloud The pieces an AI application needs, already wired to each other and reached with the same key ### The model catalog The Zen family we train ourselves, and frontier models from elsewhere. Change the model name in the request; nothing else changes. ### Agents A Python SDK where an agent is a model, instructions and tools. Put several behind a router when one prompt stops being enough. ### Inference Batching and caching happen on our side. You send a request and read a stream back; the throughput work is not yours to do. ### Vectors Embeddings stored and indexed where the rest of your data already is, so retrieval is a query rather than a second service to run. ### Evals Run a set of cases against a model, keep the results, and compare them after you change the prompt — the only way to know a change helped. ### What every call did Model, latency, tokens and cost per request, per key. A bill that surprises you is a bill you could not see coming. ### Guardrails zen3-guard classifies content before or after a turn. Scope which tools an agent may reach, and keep code execution inside a sandbox. ### Load The same endpoint answers one request a day and a sustained burst. Set a rate limit per key so one client cannot spend another one's headroom. ### SDKs Python, TypeScript, Go and Rust clients, plus a plain HTTP API for everything else. Same endpoints, same request shape. ### Your own weights Bring a model you trained and serve it behind the same API, with the same keys, limits and cost reporting as everything else in the catalog. ### Fine-tuning Start from an open-weight Zen model, train on your data, and get back a model name you can call like any other. ## One API, one key These are separate products elsewhere, each with its own account, its own key and its own bill. Here they are one endpoint that already knows who you are ### Models The Zen family and frontier models from elsewhere, all reached by name at the same endpoint ### Vectors Embedding storage and indexing, so retrieval sits next to your data instead of in a service beside it ### Documents Upload a PDF or a folder of text, get it chunked and indexed, and ask questions across all of it ### Search Keyword and meaning in one query, merged by rank, so an exact product code and a vague description both land ### Agents A model, instructions, and tools it may call — one agent, or several behind a router that picks ### Code models zen5-coder reads a million tokens of context, which is most repositories, so it answers about the code as it is ### Workflows Steps that branch, run in parallel, or loop — for work whose order matters and that a prompt cannot express ### Cost and latency Per call, per key, per model. Where the money went, and which request was slow ### Keys and secrets Sign in through Hanzo IAM, keep credentials in Hanzo KMS, and scope every key to what it is allowed to reach ### Chat A chat surface you can open, embed, or self-host — with threads, streaming, and tools already attached ### Prompts Keep versions, run them against cases, and see which one you actually shipped ### Where it runs Our regions, your cluster, or your laptop. Open weights mean the third option is real ## AI Engineering Platform Build, deploy, and scale AI applications with an integrated suite of tools designed for modern engineering teams. Hanzo AI Engineering Platform AI Models Observability Analytics ### AI Model Registry Deploy Model Import zen4LLM Provider: Hanzo Accuracy 96% Latency 85ms GPT-4oLLM Provider: OpenAI Accuracy 97% Latency 120ms ClaudeLLM Provider: Anthropic Accuracy 96% Latency 140ms Gemini 2.5LLM Provider: Google Accuracy 95% Latency 110ms Zen4 ProLLM Provider: Hanzo Accuracy 94% Latency 80ms MixtralLLM Provider: Mistral Accuracy 93% Latency 85ms Llama 4LLM Provider: Meta Accuracy 94% Latency 90ms Zen4 MiniLLM Provider: Hanzo Accuracy 95% Latency 75ms Stable DiffusionImage Provider: Stability AI Accuracy 92% Latency 200ms Cohere CommandLLM Provider: Cohere Accuracy 91% Latency 95ms ### Model Integration Integrate with OpenAI, Hugging Face, and other machine learning platforms. Deploy and serve custom ML models with built-in scaling and monitoring. - •One-click API connections to popular ML services - •Simple deployment of custom models with containerization - •Performance optimization for inference workloads ### Vector Search High-performance vector database capabilities for semantic search, RAG applications, and similarity matching across billions of vectors. - •Advanced indexing for fast k-NN and ANN queries - •Built-in embeddings generation from text and images - •Hybrid search combining vector and traditional queries ### Data Processing Specialized data processing pipelines for cleaning, transforming, and enriching training data for machine learning models. - •Automated ETL workflows for AI data preparation - •Data versioning and lineage tracking - •Scalable batch and stream processing ### AI-Enhanced Features Ready-to-use AI capabilities that can be integrated into applications with minimal configuration. - •Content generation and summarization - •Image and video analysis with computer vision - •Real-time anomaly detection and predictive analytics ## Unified Model Access Access thousands of AI models through a single, unified API with consistent interfaces and predictable pricing ### Zen — open weights from Zoo Labs Foundation Language, code, vision, image, audio, speech and retrieval, built by Zoo Labs Foundation and served on the Hanzo API. Most have their weights published, so you can serve them yourself #### Zen5 Our current frontier model, trained for agent work, reasoning step by step by default. via Zoo Labs Foundation - 1M+ context window - Agentic-trained - MoDE + CoT - Preview #### Zen5 Pro High-throughput agentic model for demanding production workloads. via Zoo Labs Foundation - 512K context window - Agentic-trained - Production optimized - Preview #### Zen5 Max Maximum context agentic model for document-scale analysis. via Zoo Labs Foundation - 2M context window - Extended CoT - Document-scale - Preview #### Zen5 Ultra Deepest reasoning model with multi-pass chain-of-thought. via Zoo Labs Foundation - 1M context window - Deep CoT - Self-verification - Preview #### Zen5 Mini Efficient agentic model with zen5-class intelligence. via Zoo Labs Foundation - 256K context window - Agentic-trained - Cost efficient - Preview #### Zen4 Flagship model for complex reasoning and multi-domain tasks. via Zoo Labs Foundation - 744B MoE (40B active) - 202K context window - Ultra Max tier - $3 / $9.60 per MTok #### Zen4 Ultra Maximum reasoning with extended chain-of-thought. via Zoo Labs Foundation - 744B MoE (40B active) + CoT - 262K context window - Ultra Max tier - Deep reasoning #### Zen4 Pro High-capability model with efficient MoE architecture. via Zoo Labs Foundation - 80B MoE (3B active) - 131K context window - Ultra tier - Efficient MoE #### Zen4 Max Most capable model for complex reasoning and agentic tasks. via Zoo Labs Foundation - Dense architecture - 1M context window - Ultra Max tier - Agentic coding #### Zen4.6 Extended context for long-document analysis and agentic workflows. via Zoo Labs Foundation - Dense architecture - 1M context window - Ultra tier - Cost efficient #### Zen4 Mini Ultra-fast lightweight model, ideal for free tier. via Zoo Labs Foundation - Dense architecture - 128K context window - Starter tier - Free tier #### Zen4 Thinking Dedicated reasoning with explicit chain-of-thought. via Zoo Labs Foundation - 80B MoE (3B active) + CoT - 131K context window - Pro Max tier - Chain-of-thought #### Zen4 Coder Code-specialized MoE for generation, review, and debugging. via Zoo Labs Foundation - 480B MoE (35B active) - 163K context window - Ultra tier - Code generation #### Zen4 Coder Pro Full-precision BF16 code model for complex codebases. via Zoo Labs Foundation - 480B Dense BF16 - 131K context window - Ultra Max tier - Full-precision #### Zen4 Coder Flash Lightweight code model for speed and inline completions. via Zoo Labs Foundation - 30B MoE (3B active) - 262K context window - Pro Max tier - Fast completions #### Zen3 Omni Multimodal model supporting text, vision, audio, and structured output. via Zoo Labs Foundation - ~200B Dense Multimodal - 202K context window - Pro Max tier - Text + Vision + Audio #### Zen3 VL Vision-language model for image understanding and visual reasoning. via Zoo Labs Foundation - 30B MoE (3B active) - 262K context window - Pro Max tier - Image understanding #### Zen3 Nano Ultra-lightweight model for edge deployment. via Zoo Labs Foundation - 8B Dense - 128K context window - Starter tier - Free tier #### Zen3 Guard Content safety classifier for moderation and guardrails. via Zoo Labs Foundation - 4B Dense - 65K context window - Pro tier - 119 languages #### Zen3 Image Best general-purpose image generation. via Zoo Labs Foundation - Diffusion - Text-to-image - Image editing - $0.04/image #### Zen3 Image Max Maximum quality image generation. via Zoo Labs Foundation - Diffusion - Maximum quality - Professional creative - $0.08/image #### Zen3 Image Dev Development model for experimentation. via Zoo Labs Foundation - Diffusion - Development - Iteration - $0.0005/step #### Zen3 Image Fast Fastest image model for real-time generation. via Zoo Labs Foundation - Diffusion - Ultra-fast - Real-time - $0.00035/step #### Zen3 Image SDXL High-resolution image generation at 1024px. via Zoo Labs Foundation - Diffusion - 1024px - High-resolution #### Zen3 Image Playground Aesthetic model for artistic generation. via Zoo Labs Foundation - Diffusion - Aesthetic - Artistic #### Zen3 Image SSD Fastest diffusion model for real-time generation. via Zoo Labs Foundation - 1B Diffusion - Fastest - Real-time #### Zen3 Image JP Japanese-specialized image generation. via Zoo Labs Foundation - Diffusion - Japanese - Specialized #### Zen3 Audio Best quality speech-to-text transcription. via Zoo Labs Foundation - 1.5B ASR - 100+ languages - Best accuracy #### Zen3 Audio Fast Fastest speech-to-text for high-throughput. via Zoo Labs Foundation - 809M ASR - Fastest - Batch optimized #### Zen3 ASR Real-time streaming speech recognition. via Zoo Labs Foundation - Streaming ASR - Real-time - Sub-500ms latency #### Zen3 ASR v1 First-generation streaming ASR. via Zoo Labs Foundation - Streaming ASR - Legacy - Compatible #### Zen3 TTS High-quality text-to-speech with natural prosody. via Zoo Labs Foundation - 82M TTS - 40+ voices - 8 languages #### Zen3 TTS HD Maximum fidelity text-to-speech. via Zoo Labs Foundation - TTS HD - Broadcast-grade - 48kHz output #### Zen3 TTS Fast Low-latency TTS for real-time voice agents. via Zoo Labs Foundation - 82M TTS - Low latency - Voice agents #### Zen3 Embedding High-quality text embeddings for RAG and search. via Zoo Labs Foundation - 3072 dimensions - 8K context window - Pro Max tier #### Zen3 Embedding Medium Balanced embedding model for retrieval. via Zoo Labs Foundation - 4B parameters - 40K context window - Cost-effective #### Zen3 Embedding Small Lightweight embedding for high throughput. via Zoo Labs Foundation - 0.6B parameters - 32K context window - High-throughput #### Zen3 Embedding OpenAI OpenAI-compatible embedding endpoint. via Zoo Labs Foundation - 3072 dimensions - 8K context window - OpenAI compatible #### Zen3 Reranker High-quality reranker for RAG pipelines. via Zoo Labs Foundation - 8B parameters - 40K context window - RAG accuracy #### Zen3 Reranker Medium Balanced reranker for retrieval. via Zoo Labs Foundation - 4B parameters - 40K context window - Cost-effective #### Zen3 Reranker Small Lightweight reranker for high throughput. via Zoo Labs Foundation - 0.6B parameters - 40K context window - Minimal cost ### Models from other labs Same endpoint, same key, same request shape. Reaching for one of these is changing a string, not a migration #### Claude Opus 4.6 Anthropic's top tier, for the questions where depth beats speed. via Anthropic - 1M context window - Complex reasoning - Extended thinking #### Claude Sonnet 4.6 Ideal balance of capability and speed for production workloads. via Anthropic - 1M context window - Best balance of speed and intelligence - Strong coding - Fast inference #### Claude Haiku 4.5 Fastest and most affordable Claude model for high-throughput tasks. via Anthropic - 200K context window - Fastest Anthropic model - Cost efficient - Low latency #### GPT-4o OpenAI's flagship model with advanced reasoning capabilities. via OpenAI - 128K context window - Multimodal - Advanced reasoning - Tool use #### GPT-4o Mini Cost-efficient OpenAI model for everyday tasks. via OpenAI - 128K context window - Fast and affordable - Good quality - Low latency #### Gemini 3.1 Pro Google's flagship with the longest context window. via Google - 1M context window - Multimodal - Long-context reasoning - Code generation #### GLM-5.2 The newest GLM on the gateway. Reach for it the way you reach for any other model here — change the string. via Z.ai - Agentic tool use - Strong coding - Served as z-ai/glm-5.2 #### GLM-4.7 The previous GLM generation, still a common default for high-volume work. via Z.ai - Cost efficient - High throughput - Served as z-ai/glm-4.7 ### Custom Models Deploy and customize models to meet your specific needs #### Fine-tuned Models via Custom - Domain adaptation - Company knowledge base - Specialized tasks - Improved performance #### Hugging Face Models via Custom - Community models - Thousands of options - Specialized capabilities - Open source #### Custom Embedding Models via Custom - Domain-specific embeddings - Custom similarity metrics - Enhanced search - Optimized retrieval ### All of it at one address Routing, documents, vectors and knowledge bases answer at api.hanzo.ai with the key you already have #### Routing Ask for auto and Enso classifies the turn, prices it, and picks — so an easy question does not pay frontier rates #### Documents Parsing and chunking happen here, so a PDF becomes searchable text without a pipeline you maintain #### Vectors Embeddings stored and indexed next to your data, so retrieval is a query rather than a second service to run #### Knowledge bases Group documents into a base, point a model at it by name, and every answer carries the passages it came from Explore the API ## What people build with it Six shapes that keep coming back, and the models each one reaches for ### Support that reads your docs An assistant grounded in your own content, answering from what you published rather than from what it half-remembers. ### Text, images, audio, video zen-image draws, zen-voice speaks, zen-music composes, zen-video moves. Same key, same billing, one request each. ### Search over your own data Index what you have, then answer questions from it with the passages that support the answer attached. ### Work that runs without you An agent with tools, a trigger, and a place to write the result. Triage a queue, reconcile a ledger, file the report. ### Coding agents Hanzo Dev in the terminal, the MCP tool surface underneath it, and a code model that fits the whole repository in context. ### Speech both directions zen3-asr transcribes thirty languages and can stream while someone is still talking. zen3-tts speaks back. ### What it looks like A client, a model name, and a message. The vector store and the agent below use the same client and the same key ``` import { Hanzo } from '@hanzo/ai'; // Initialize the Hanzo AI client const hanzo = new Hanzo({ apiKey: process.env.HANZO_API_KEY }); // Create a conversation with memory const conversation = hanzo.conversation({ model: 'zen5', memory: true, system: 'You are a helpful assistant' }); // Send a message and get a response const response = await conversation.send('Tell me about AI engineering'); console.log(response); ``` #### Documentation Example ##### Vector Search ``` // Create a vector store const vectorStore = hanzo.vectorStore('my-store'); // Add documents to the store await vectorStore.addDocuments([ { text: 'AI engineering best practices...' }, { text: 'Deploying models to production...' } ]); // Search for similar documents const results = await vectorStore.search( 'How to deploy AI models?', { limit: 3 } ); ``` ##### AI Agents ``` // Create an agent with tools const agent = hanzo.agent({ model: 'zen5-max', tools: [ hanzo.tools.webSearch(), hanzo.tools.codeInterpreter(), vectorStore.asTool('knowledge') ] }); // Run the agent with a task const result = await agent.run( 'Analyze our production metrics and suggest optimizations' ); ``` ## Experiences from Our Community Hear from engineering teams who are building the next generation of AI-powered applications "Hanzo gave us the infrastructure backbone to move fast without rebuilding from scratch. The platform let our team focus on the product, not the plumbing." JG Jay Giraud CEO, Damon Motorcycles "We needed a platform that could handle real-time data at scale without sacrificing developer experience. Hanzo delivered on both fronts." MW Marcus Weller CEO, SKULLY Technologies "Hanzo's AI infrastructure helped us personalize experiences for millions of users while keeping our stack lean and our team focused on what matters." SM Sandro Mur CEO, Bellabeat ### The AI Engineering Community Build with the engineers shipping production AI. Share experiences, get support, and collaborate on best practices. Active developer community Weekly office hours Dedicated support team JS MK AL TN RW + ## Make the first call Create an account, take a key, and send one request. Everything on this page answers at the same endpoint with that key Get Started Free Schedule a Demo ### Documentation Every endpoint, every parameter, and the SDK method that calls it, in four languages. [Explore Docs](#) ### Quickstart A key, a curl, and a response — then the same thing again from a real project you can clone. [Try Quickstart](#) ### Community Most of this is open source. Read the code, file the issue, send the patch. [Join Community](#) ### Still deciding? An account is free, and some models cost nothing to call. Try it before you talk to anyone. Create Free Account Contact Sales