Hanzo 部落格
前沿 AI 模型、代理基礎架構，以及驅動下一代智慧應用的平台。
代理型語言模型的時代，無頭式商務要先賣得成
買家正在變成代理人。要讓大型語言模型在沒有人插手時讀得懂、簽得了名、付得了款，需要哪些條件，以及 Hanzo 今天為此準備了什麼。
你的 AI 正在盤算取你性命
最前線模型面對真實人命攸關時為何失靈，以及 Hanzo Kai 1.9 如何以確定性政策落實安全。
Kai 1：決策，而不是文字補全
Kai 是 Hanzo 的決策模型，透過 Hanzo API 提供。給它明確的狀態，以及列出選項的問題，它會回傳具型別、經過校準的答案——一項選擇、一個分數、一個是／否，或一次有理由的延後。
什麼是決策模型？
生成模型在不受約束的段落裡進行推理。決策模型則以具型別、經過校準的狀態轉移來做決定。這就是為什麼人工智慧的下一個前沿並非生成，而是決策。
Hanzo AI 解開的投資機構關鍵痛點
找出私募股權、創業投資與資產管理裡真正卡住營運的難題，並看 Hanzo 的主權人工智慧雲端、前沿推理模型與代理工作流如何把案源開發、盡職調查與執行做順。
路由器就是登錄表
Hanzo Cloud 沒有簽入的 OpenAPI 檔案。活著的路由表才是源頭，朝三個方向投影，任一投影一漂移，雙射測試就讓建置失敗。963 條路徑上的 1,362 個操作，在請求當下按行程產生。
GB10 上每秒五十萬請求：瓶頸是網路，不是那台機器
我們想知道雲端資料平面在 NVIDIA GB10 上能跑多快。得到的教訓比較像網卡接收佇列，不太像我們自己的程式碼——同一條線上，吞吐量還是超過原本的兩倍。
即日起正式推出 Hanzo Datastore：迎向後量子時代的欄式 OLAP 分析引擎
這是從 ClickHouse 分出的欄式倉儲，依 Hanzo S3 上儲存與運算分離的方式重做。透過剖析導引與連結期最佳化，常駐記憶體大約只剩一半，gRPC 也已移除，並留下一條通往無領導者、後量子協調的路。
在同一個行程裡終結 Raft：改以無領導者的後量子共識 Lux Quasar 換下 ZooKeeper 的共識引擎
本文說明 Hanzo Datastore 如何把 ClickHouse Keeper 底下的 NuRaft 引擎換成 Lux Quasar。換上的是後量子、沒有單一領導者的共識，直接連進同一套 C++ 伺服器，複製層繼續呼叫原來的 ZooKeeper 介面，呼叫方式維持不變。
人工智慧證明：付錢給它所執行的開源開發者的雲端
Hanzo Cloud 站在人工智慧證明上——任何驗證者都能在鏈上重驗的確定性推論——也站在一項承諾上：雲端賺到的 25%，流回那些程式碼真正被它執行的開源開發者。
Hanzo Brain：面向代理的本機優先知識圖譜
一個 SQLite 檔案，你的代理從每個執行環境讀取它。圖操作不做語言模型往返，寫入時同時做 FTS5、向量和帶型別邊的抽取。
Zen：1T+ 參數的混合專家架構
Hanzo 的旗艦語言模型：由頂尖語言模型融合而成的 1.04 兆參數混合專家模型，帶有分散式訓練與鏈上模型證明。
Hanzo 發表 AI 代理基礎架構的形式化證明
針對代理安全、MCP 協定、PoAI 共識與平台營運、由機器檢驗的 Lean 4 證明
BitDelta：從單一圖形處理器叢集提供 14 個 Zen 模型
BitDelta 以 1 位元增量壓縮，讓我們從共用的圖形處理器基礎建設提供 14 個 Zen 模型變體——數學、架構與取捨。
從商業智慧到前沿模型：Hanzo 十年來的基礎架構
打造 AI 基礎架構的 12 年。自 2014 年的即時推薦引擎起，到 2026 年的 94+ 前沿模型與後量子區塊鏈。這是我們走到這裡的技術紀錄。
完整的 AI 代理程式堆疊：模型、運算與工具，放在同一個平台
Hanzo AI 推出第一個把 100 多個 AI 模型、雲端運算、GPU 存取，以及 260 多個 MCP 工具放進同一個開發者帳戶的平台。
按每小時 $3.48 提供 H100 顯卡，比 AWS 便宜 72%
Hanzo Bot 以每小時 $3.48，向 AI 代理程式開發者開放隨需 H100 GPU，比 AWS 的同等價格低 72%。
每月 $5 起，把模型、運算和工具放進同一個 AI 代理程式平台：Hanzo Bot
Hanzo Bot 以業內最便宜的 AI 代理程式託管上線——從每月 $5 起，含免費額度，整合存取 100 多個模型，以及每小時 $3.48 的 H100 GPU。
AI 轉錄比 OpenAI 便宜 85%：給機器人開發者的多模態價格
Hanzo Bot 以每分鐘 $0.0009 提供語音轉文字——比 OpenAI Whisper 便宜 85%——影像產生則是業界最低的每步 $0.00013。
Hanzo 定價深度解析：每一種模型、每一種方案、每一次比較
完整拆解 Hanzo 在 AI 模型、雲端運算、GPU 存取與工具上的定價，並與每一家主要對手做正面比較。
一套介面通往每一款 AI 模型：Hanzo AI 閘道正式推出
Hanzo AI 把業界第一個不加價、串起多家供應商的 AI 閘道送上線：一把 API 金鑰就能用到各大供應商的 100+ 模型，再加 14 個自家的 Zen 模型。
Zen4 Coder：480B 參數裡，真正動用的少了 92%，面向程式碼的專家混合模型
Hanzo AI 推出 Zen4 Coder，這是 480B 參數的 Mixture of Experts 程式碼模型，每個 token 只啟動 35B 參數——以運算成本的一小部分，帶來前沿的程式碼智能。
零出站費：該退場的雲端定價慣例
出站費是對成功課稅。Hanzo Network 在每一檔都包含頻寬——從 $5 方案的 500 GB 到 Ultra 的 120 TB——沒有超量費用。
現已上線並可呼叫：高效率混合專家與前沿推論模型 Qwen3.5-35B-A3B 與 Mercury 2
Qwen3.5-35B-A3B（Apache 2.0，多模態 MoE）與 Inception 的 Mercury 2 現已在 Hanzo AI 閘道提供。兩者都透過同一金鑰呼叫，不必另開帳戶，也不必換一套計費。
現已上線並可在閘道使用：Google 的 1M 詞元多模態推論模型 Gemini 3.1 Pro
Google 的 Gemini 3.1 Pro 現已在 Hanzo AI 閘道提供——具備 1M 詞元上下文視窗，以及橫跨文字、影像、音訊、影片與程式碼的多模態推論。
現已上線可供呼叫：大規模多模態混合專家模型 Claude Sonnet 4.6 與 Qwen3.5-397B
Claude Sonnet 4.6 與 Qwen3.5-397B-A17B 現已在 Hanzo AI 閘道提供——Anthropic 的最新模型，以及最大的開源多模態 MoE 模型。兩者都透過同一金鑰呼叫，不必另開帳戶。
現已上線可供使用的兩個新模型：前沿多模態推論與開源程式碼代理 ByteDance Seed 2.0 與 Qwen3-Coder-Next
ByteDance 的 Seed 2.0 Pro 與阿里巴巴的 Qwen3-Coder-Next 現已在 Hanzo AI 閘道提供——前沿多模態推論與開源程式碼代理。兩者都已接入同一閘道，無需另開帳戶。
現已上線：參數規模空前的兩個開源巨頭 GLM-5（744B）與 MiniMax M2.5（230B）
GLM-5 與 MiniMax M2.5——有史以來發布的最大開源模型中的兩個——現已透過 Hanzo AI 閘道提供。
現已上線可供使用：面向裝置端的多模態人工智慧，Google Gemma 3n 與史上最小的 Gemma 3 270M
Google 的 Gemma 3n（裝置端多模態）與 Gemma 3 270M（史上最小的 Gemma）現已在 Hanzo AI 閘道提供。兩者都適合在手機和邊緣裝置上執行。
現已上線並可立即呼叫：零加價的同一介面與同一金鑰，Hanzo 閘道上的旗艦模型 Claude Opus 4.6 與程式碼模型 GPT-5.3 Codex
Anthropic 的 Claude Opus 4.6 與 OpenAI 的 GPT-5.3 Codex 現已透過 Hanzo AI 閘道提供——同一 API、同一金鑰、零加價。兩者都在發布後不久即可從同一入口呼叫。
Hanzo Cap Table：給新創的股權管理
推出 Hanzo Cap Table。用它管理股權、試算融資情境，並為四百零九Ａ評估做好準備，而不必支付 Carta 那樣的年費。
英文
尚未翻譯：這些文章會以英文開啟。
Introducing Enso — one API, routed to the best frontier intelligence
Enso is Hanzo's learned router: no single frontier model leads every benchmark, so Enso routes each request to the model most likely to win it — and learns from your feedback as it goes. Microsecond routing on a CPU, a transparent per-request meter, and a 1% fee that pays for itself. Here is how it works, the measured results, and the cost-transparency angle a monolith cannot offer.
Route every prompt to the right model — and cut AI cloud costs by up to 90%
Hanzo Router places each request on the cheapest model that can still serve it — reuse a local model for free, load one that fits, or fall through to cloud. Here's the idea, what it feels like in production (qualitative), and the honest cost arithmetic (quantitative) — no fabricated benchmark.
Enterprise Cloud Without Enterprise Pricing: 96 vCPUs at $3,999/Month
Hanzo Network launches Business through Ultra enterprise cloud tiers — dedicated CPU machines from 8 to 96 vCPUs with 71–81% cost advantage over hyperscaler equivalents.
Hanzo Network: 10–54% Cheaper Than DigitalOcean, Lightsail, and Vultr
Hanzo Network launches a developer cloud that undercuts DigitalOcean, AWS Lightsail, Vultr, and Linode by 10–54% — with zero egress fees, DDoS protection, and consistent global pricing included.
Zen4: Abliterated AI Models for Every Scale
Announcing the full Zen4 family: mini (4B) through ultra (1T MoE), all abliterated. Eight models covering every scale from edge to cloud, with no refusal behavior and full capability access.
Introducing Zen4: Open Foundation Models from 4B to 1T+
Zen4 is a complete lineup of open AI models spanning from 4B to over 1 trillion parameters, featuring consumer, coder, and ultra tiers.
Hanzo Browser Extension: AI in Every Tab
Bringing AI capabilities to your browser - summarize pages, answer questions, and automate workflows.
Hanzo Dataroom: Secure Document Sharing with Analytics
Introducing Hanzo Dataroom — secure document sharing with per-page analytics, watermarking, and granular access controls for fundraising and M&A.
Hanzo ML Platform: End-to-End Machine Learning Infrastructure
Announcing the unified Hanzo ML Platform — training, serving, pipelines, feature store, model registry, and evaluation in one integrated system.
Zen Reranker: Neural Reranking at 30x Compression
Zen Reranker delivers 7680-dimensional neural reranking with a 31.87x BitDelta compression ratio, designed for RAG pipelines and integrated with Hanzo and Zoo decentralized search networks.
Hanzo Dev: AI Coding Agent for Your Terminal
Hanzo Dev launches — an AI coding agent that lives in your terminal, understands your codebase, and writes production-quality code.
Zen VL: Vision-Language Models with Function Calling
Zen VL is a family of vision-language models at 4B, 8B, and 30B -- each with instruct and agent variants -- supporting OCR in 32 languages, GUI navigation, spatial grounding, and native function calling with visual context.
Zen Artist: Image Generation Reimagined
Introducing Zen Artist - our multimodal image generation model with unprecedented control and quality.
Zen Omni: Unified Multimodal AI
Zen Omni is a 30B MoE unified multimodal model with Thinker-Talker architecture, handling text, vision, and audio in a single model with real-time speech-to-speech at under 300ms latency.
MCP: 260 Tools for AI Models
How we built 260+ tools using Model Context Protocol, enabling AI models to interact with the world.
Zen Designer: 235B Vision-Language Model
Zen Designer is a 235B MoE vision-language model with 22B active parameters, supporting image analysis, video understanding, OCR in 32 languages, and native design reasoning.
Zen Max: 671B Reasoning Model
Zen Max is a 671B MoE reasoning model with 384 experts, 256K context, and abliterated base weights -- achieving AIME 2025 99.1%, SWE-Bench 71.3%, and BrowseComp 60.2%.
One Billion Dollars in Client Revenue
Hanzo AI has driven over $1 billion in cumulative revenue for our clients through AI-powered marketing and commerce infrastructure.
Zen Coder: Code AI from 4B to 480B
Zen Coder is a family of code-specialized models spanning 4B edge to 480B frontier, with 128K context, extended thinking up to 512K tokens, MCP integration, and support for 100+ programming languages.
Hanzo Operative: Computer Use for AI Agents
Introducing Hanzo Operative — enabling AI agents to use computers like humans: screen capture, mouse, keyboard, and application interaction.
Hanzo Flow: Visual AI Workflow Builder
Hanzo Flow launches — a visual drag-and-drop builder for complex AI workflows. Connect models, tools, and APIs without writing integration code.
Zen Guard: Multilingual AI Safety Classification
Zen Guard is an 8B multilingual safety classifier covering 119 languages and 9 harm categories, with a three-tier severity system and 5ms/token streaming latency.
Introducing Hanzo Console: Unified LLM Observability
Hanzo Console launches — a unified development environment for observing, debugging, and optimizing LLM applications across all providers.
Zen Embedding: #1 Multilingual Retrieval
Zen Embedding reaches #1 on the MTEB multilingual leaderboard with a 70.58 score, supporting 100+ languages and Matryoshka representation learning across four dimension sizes.
Hanzo Experiments: Feature Flags and A/B Testing
Introducing Hanzo Experiments — feature flags, A/B testing, and progressive rollouts with statistical rigor.
Zen Nano: 0.6B Edge AI Model
Zen Nano is a 0.6B on-device AI model built for mobile and embedded deployment, achieving 44K tokens/sec on M3 Max with a 40K context window.
Zen Musician: AI-Generated Music That Sounds Human
Introducing Zen Musician - our AI music generation model that creates original compositions in any style.
Jin 2.0: The Next Generation of Multimodal AI
Announcing Jin 2.0, a major upgrade to our unified multimodal AI architecture with expanded capabilities and improved performance.
Case Study: Keek — AI-Powered Social Media Transformation
How Hanzo integrated AI infrastructure into the Keek social platform — content recommendation, multimodal moderation, and creator tools processing 12M daily content items.
ACI: The AI Compute Infrastructure Network
Introducing ACI, a decentralized network for AI compute that enables trustless, verifiable AI inference.
Zen Video: Generative Video at Scale
Introducing Zen Video - our generative video model for creating high-quality video content from text and images.
Hanzo Partners with Personas Social to Power Keek's AI
Hanzo AI partners with Personas Social Inc. to integrate AI-powered features into the Keek social platform.
Hanzo Notebooks: Managed Jupyter with GPU Access
Introducing Hanzo Notebooks — managed Jupyter workspaces with on-demand GPU access and pre-installed ML frameworks.
Candle: Rust-Native Machine Learning
Introducing our investment in Candle, the Rust ML framework powering Hanzo's next-generation inference.
Hanzo Sign: Electronic Signatures Without the Enterprise Tax
Introducing Hanzo Sign — legally binding electronic signatures with audit trails, templates, and API access at a fraction of DocuSign's price.
Hanzo CX: Complete Customer Experience Platform
Introducing Hanzo CX — a unified customer experience platform with omnichannel inbox, AI-assisted responses, and full conversation history.
Zen Coder: Code Generation That Actually Works
Building Zen Coder - our code generation model trained on real software engineering workflows.
Zen Dub: Voice Cloning and AI Dubbing
How we built Zen Dub - an AI system for voice cloning and automatic dubbing in 100+ languages.
LLM Infrastructure: Running AI at Scale
How we built infrastructure to serve billions of LLM requests for commerce applications.
Hanzo Auto: AI Workflow Automation
Introducing Hanzo Auto — AI-powered workflow automation with 280+ integrations and a visual builder.
Hanzo CRM: Modern Customer Relationship Management
Introducing Hanzo CRM — a modern, AI-enhanced CRM for teams that want to close deals, not manage software.
Zen Agent: Building AI That Actually Does Things
Introducing Zen Agent - our framework for building AI systems that can take action in the real world.
Zen Integration: AI That Follows Your Rules
Introducing Zen, our system for ensuring AI follows business rules, compliance requirements, and brand guidelines.
Hanzo Insights: Product Analytics That Respects Privacy
Introducing Hanzo Insights — product analytics with event tracking, funnels, cohorts, and session replay, without selling user data.
Hanzo Edge: High-Performance API Gateway
Introducing Hanzo Edge — a high-performance API gateway with sub-millisecond routing, rate limiting, and authentication at the edge.
Gym: Distributed Training for the Rest of Us
Introducing Gym - our platform for distributed AI model training that turns the world's GPUs into a supercomputer.
Jin: Unified Multimodal AI Architecture
Introducing Jin, our unified architecture for multimodal AI that powers the next generation of Hanzo intelligence.
LLM Inference at the Edge: Our llama.cpp Journey
How we brought large language models to edge devices using llama.cpp and what we learned along the way.
Hanzo Observability: Metrics, Logs, and Traces in One Platform
Launching the Hanzo Observability stack — unified metrics, logs, and distributed tracing with OpenTelemetry-native ingestion.
hanzo.vim: AI-Powered Vim for Terminal Purists
Bringing AI capabilities to Vim without sacrificing the terminal-native experience.
Identity NFTs: Portable Commerce Identity
Introducing Identity NFTs: blockchain-based portable identity for commerce.
Open Source at Hanzo: Building in Public
Announcing Hanzo's open-source strategy — why we're open-sourcing our core infrastructure and introducing revenue sharing for contributors.
HKE: Hanzo Kubernetes Engine
Launching HKE — managed Kubernetes that eliminates the operational burden without limiting what you can do.
Building Our ML Framework
Why we built a custom ML framework and what we learned along the way.
Case Study: Bellabeat — Scaling a Women's Health Platform
How Hanzo helped Bellabeat scale their women's health wearable platform to $35M in projected revenue through CMO/CTO services, marketing automation, and data analytics.
Hanzo Search: Full-Text Search That Just Works
Introducing Hanzo Search — full-text search with typo tolerance, faceting, and sub-millisecond query times.
Agent SDK: Build Intelligent Assistants
Introducing the Hanzo Agent SDK: tools for building intelligent commerce assistants.
Hanzo Platform: Git Push to Production
Launching Hanzo Platform — a PaaS that deploys your code from git push to production with zero DevOps configuration.
Multi-Modal AI for Commerce
How we are combining vision, language, and structured data for next-generation commerce AI.
Hanzo Tasks and Queues: Durable Background Execution
Launching Hanzo Tasks and Queues — durable workflow execution and message queues for reliable background processing.
Case Study: Damon Motorcycles — 500x ROI Campaign
How Hanzo delivered a 500x ROI marketing campaign for Damon Motorcycles, generating a 230% increase in qualified leads for their electric motorcycles.
The Context Layer: Our Early Work on Multi-Model State Management
How we built a context management protocol for multi-model AI pipelines — early research that shaped how we think about AI state.
Hanzo KMS: Secrets Management Done Right
Introducing Hanzo KMS — end-to-end encrypted secrets management with environment sync, rotation, and audit trails.
Hanzo Storage: S3-Compatible Object Storage
Launching Hanzo Storage — S3-compatible object storage with integrated CDN and no egress fees.
Case Study: Casper — Enterprise Blockchain from Concept to Mainnet
How Hanzo helped Casper Labs architect and launch an enterprise-grade blockchain — from the Rust pivot to mainnet launch and DEVxDAO founding.
Hanzo SQL: Managed PostgreSQL with Zero Ops
Launching Hanzo SQL — managed PostgreSQL with automatic scaling, high availability, and zero operational overhead.
Introducing the Hanzo Agent Framework
Announcing the Hanzo Agent Framework: building blocks for intelligent automation in commerce.
Hanzo Base: Backend-as-a-Service for Modern Applications
Introducing Hanzo Base — a complete backend-as-a-service with database, auth, storage, and realtime capabilities.
Hanzo Cloud: Managed Infrastructure for Commerce
Introducing Hanzo Cloud: fully managed commerce infrastructure with enterprise-grade reliability.
Hanzo DNS and Networking: Programmable Infrastructure
Launching Hanzo DNS and Networking — programmable DNS with service discovery and zero-trust virtual networks.
Webhooks: Real-Time Event Delivery
Announcing Hanzo Webhooks: reliable, secure, real-time event delivery for your integrations.
Case Study: Unikoin Gold — 120,000 ETH Token Launch
How Hanzo helped Unikrn raise 120,000 ETH in their token launch, combining blockchain infrastructure with regulatory compliance across 15+ jurisdictions.
The Hanzo SDK Ecosystem
Announcing official SDKs for JavaScript, Python, Ruby, Go, and PHP, plus the new SDK development kit.
Hanzo Functions: Serverless Compute with Zero Cold Starts
Introducing Hanzo Functions — serverless compute with zero cold starts, built for event-driven commerce workloads.
The Hanzo API Gateway
Introducing the Hanzo API Gateway: unified access, better security, and improved developer experience.
Co-Founding the First SEC-Approved Crowdfunding Token
Hanzo co-founded the first SEC-approved crowdfunding token offering, bridging traditional crowdfunding with tokenized securities.
ML-Powered Analytics: From Dashboards to Decisions
How we are using machine learning to transform analytics from passive reporting to active decision support.
Hanzo Payments: Unified Payment Processing for Global Commerce
Launching Hanzo Payments — unified payment processing with intelligent routing, fraud detection, and multi-currency support.
Checkout 2.0: Conversion Through Simplicity
Announcing Checkout 2.0, our redesigned checkout experience built for maximum conversion.
Hanzo Identity: Authentication That Scales with Your Business
Introducing Hanzo Identity — enterprise-grade authentication and access management built for commerce platforms.
Commerce AI: What We Learned in 2017
Lessons from a year of building and deploying AI systems for commerce.
42 Million in Client Sales: What Techstars Taught Us
Reflecting on Techstars Demo Day and $42M in client sales generated through the Hanzo platform.
Hanzo Joins Techstars '17
Hanzo has been selected for Techstars 2017, accelerating our mission to power intelligent commerce.
Introducing the Hanzo Platform
Crowdstart evolves into Hanzo: a complete commerce platform for the modern era.
AI-Powered Product Recommendations
How we built collaborative filtering and content-based recommendations into the Crowdstart platform.
Hanzo Analytics: Real-Time Intelligence for Commerce
Launching Hanzo Analytics — real-time behavioral intelligence for commerce teams who need answers now, not tomorrow.
Earle: Genetic Algorithms for Marketing Optimization
Introducing Earle, our genetic algorithm system for evolving high-performing marketing campaigns.
Real-Time Analytics: From Data to Decisions
How we built real-time analytics into Crowdstart and what we learned about data-driven commerce.
Astle.js: Reactive Commerce Components
Releasing Astle.js, our open-source library for building reactive commerce interfaces.
Cloud-Native Commerce: Why Architecture Matters
How we built Crowdstart as a cloud-native platform from day one, and why this matters for commerce.
Introducing Crowdstart: Commerce for the Crowdfunding Era
Announcing Crowdstart, a new platform enabling creators to build sustainable businesses from successful crowdfunding campaigns.
PubSub: The Event-Driven Foundation
How we built a publish-subscribe system that became the nervous system of our entire platform.
KV: Distributed State for the Real World
Building a key-value store that could handle the chaos of distributed systems - our second foundational project.
The Datastore: Where It All Began
The first building block of what would become Hanzo - a flexible datastore designed for the next generation of applications.