# Hanzo Platform — one stack, down to the kernel > Models, agents, applications, data and infrastructure on one operating plane, from GPU kernel to autonomous agent. Run it managed or on infrastructure you control. Hanzo Platform # The AI‑native cloud. Run models, agents, applications, data and infrastructure through one operating plane. Kubernetes-native. Fully observable. Open source. Run it on Hanzo infrastructure or infrastructure you control. [Start building](https://platform.hanzo.ai/api-keys)[Inspect the stack](#stack) 500+ callable models · GPUs · Functions · Containers · Databases · Observability · One API $ docker run -p 8080:8080 ghcr.io/hanzoai/cloud Console — Hanzo [Chat](https://hanzo.ai/cloud/chat)[Bot](https://hanzo.ai/bot)[Search](https://hanzo.ai/search)[Crawl](https://hanzo.ai/crawl)[Studio](https://hanzo.ai/studio)[Console](https://hanzo.ai/console)[Referrals](https://hanzo.ai/cloud/referrals)[Marketplace](https://hanzo.ai/cloud/marketplace) - gatewayReady - payments-apiReady - checkout-webBuilding - docs-siteReady Every request enters here, so the limiter does too. [Open in the console](https://hanzo.ai/home) Example project. Every name in it is one you would give your own. The native AI stack ## One stack.Down to the kernel. Build with agents. Route with Enso. Reason with Zen. Decide with Kai. Run on infrastructure you control. Inspect every layer Hanzo / architecture mapSelect a layer 01 Gateway 02 AI API 03 Enso 04 Zen 05 Kai 06 Agents 07 Sandboxes 08 Cloud 09 Node 10 Engine 11 Hanzo ML 12 Kernel DSL 01—12 / owned and connectedInteractive drawing · not telemetry Layer inspector03 / 12 Unified entry ### Gateway One entry for the AI stack: keys, orgs, admission, usage limits and routing to the right model or service. [Product](https://hanzo.ai/gateway)[Docs](https://docs.hanzo.ai/docs/openapi/gateway)[Source](https://github.com/hanzoai/gateway) Why this layer mattersA stable API boundary, independent of which backend runs the request. Model interface ### AI API A compatible inference interface for chat, completions, embeddings, multimodal models and agent workloads. [Product](https://hanzo.ai/api)[Docs](https://docs.hanzo.ai/docs/openapi/ai)[Source](https://github.com/hanzoai/ai) Why this layer mattersThe same client and key can move between models without rebuilding the application. Intelligent routing ### Enso Choose models and execution paths by capability, quality, latency and cost. Invoke a specific model when the application needs exact control. [Product](https://hanzo.ai/enso)[Docs](https://docs.hanzo.ai/docs/models/enso)[Examples](https://github.com/hanzoai/enso) Why this layer mattersEnso routes. Zen reasons. Kai decides. 95.6% on GPQA DiamondHistorical multi-model research evaluation · new questionsEnso Ultra answered 152 of 159 new questions right, against 93.1% for the best single model on the same questions. [How Enso routes →](https://hanzo.ai/enso) [View the measurements →](https://hanzo.ai/benchmarks/gpqa) Generative reasoning ### Zen Hanzo’s open-weight generative model family for reasoning, coding, language and multimodal generation. [Product](https://hanzo.ai/zen)[Docs](https://docs.hanzo.ai/docs/models/zen)[Model recipes](https://github.com/hanzoai/zen) Why this layer mattersThe reasoning layer: language, synthesis, plans and generation. [View the measurements →](https://hanzo.ai/benchmarks/zen-6) Decision intelligence ### Kai Typed decisions and policy-aware evaluations for choosing actions, scoring evidence and protecting agent execution boundaries. [Product](https://hanzo.ai/kai)[Docs](https://docs.hanzo.ai/docs/models/kai)[Examples](https://github.com/hanzoai/kai) Why this layer mattersReasoning can generate proposals; Kai helps turn evidence into accountable decisions. [View the measurements →](https://hanzo.ai/benchmarks#kai) Autonomous action ### Agents Composable agents coordinate tasks, use tools, schedule work and continue from intention to verified results. [Product](https://hanzo.ai/agents)[Docs](https://docs.hanzo.ai/docs/openapi/agent)[Source](https://github.com/hanzoai/agents) Why this layer mattersAgents act across software and services instead of stopping at a response. 3.02× faster codingHanzo Dev beside Claude Code on the same coding repairs, paired over 42 tasks both solved; 95% interval 2.06–4.36×. [Hanzo Dev →](https://hanzo.ai/dev) [View the measurements →](https://hanzo.ai/benchmarks#coding-agent) Safe execution ### Sandboxes Isolated execution for code, browser work and agent tools, with controlled permissions and lifecycle. [Product](https://hanzo.ai/platform)[Docs](https://docs.hanzo.ai/docs/openapi/sandbox)[Source](https://github.com/hanzoai/vm) Why this layer mattersKeep tool execution separated from model inference and production resources. Operating plane ### Cloud Deploy and coordinate inference and agent systems across Hanzo infrastructure, your clusters or a hybrid environment. [Product](https://hanzo.ai/platform)[Docs](https://docs.hanzo.ai/docs/network)[Source](https://github.com/hanzoai/cloud) Why this layer mattersOne operating plane connects identity, usage, execution and infrastructure. Execution fabric ### Node Connect your machines to the agent execution fabric, advertise capabilities and receive authorized invocations. [Product](https://hanzo.ai/node)[Docs](https://docs.hanzo.ai/docs/cli/node)[Source](https://github.com/hanzoai/node) Why this layer mattersConnected nodes can be invoked for the work they advertise. Inference runtime ### Engine The native serving engine with continuous batching, paged attention, quantization and persistent KV caching. [Product](https://hanzo.ai/engine)[Docs](https://docs.hanzo.ai/docs/openapi/engine)[Source](https://github.com/hanzoai/engine) Why this layer mattersThe inference work runs in one serving binary, from local hardware to GPU-backed systems. [View the measurements →](https://hanzo.ai/benchmarks/inference) Tensor and neural framework ### Hanzo ML Rust tensor, neural-network and model-execution primitives with accelerated device backends. [Product](https://hanzo.ai/open-source)[Docs](https://github.com/hanzoai/ml/blob/main/README.md)[Source](https://github.com/hanzoai/ml) Why this layer mattersHanzo owns framework code close to the math, not just the API wrapper. Hardware acceleration ### Kernel DSL A first-party Rust kernel DSL targeting CUDA, ROCm, Vulkan and Metal alongside backend-specific implementations. [Product](https://hanzo.ai/open-source)[Docs](https://github.com/hanzoai/ml/blob/main/hanzo-kernel/README.md)[Kernel source](https://github.com/hanzoai/ml/tree/main/hanzo-kernel) Why this layer mattersOne portable kernel expression, optimized and validated for each device. Backend performance is workload-specific. Select / exploreProduct · docs · source ZAP / native interconnectAgent ↔ tool RPC · MCP bridge · typed service communication [Protocol](https://docs.hanzo.ai/docs/zap)[Source](https://github.com/hanzoai/zap) ### One kernel DSL. Four targets. Four parallel acceleration backends, not four extra hops. First-party kernels and performance tuning for supported workloads. 01 / NVIDIACUDACUDA kernels, attention and accelerated GPU operations. [Explore CUDA](https://github.com/hanzoai/ml/tree/main/hanzo-kernels) 02 / AMDROCmHIP/ROCm execution, attention and quantized operations. [Explore ROCm](https://github.com/hanzoai/ml/tree/main/hanzo-rocm-kernels) 03 / Cross-GPUVulkanSPIR-V shader paths and portable GPU execution. [Explore Vulkan](https://github.com/hanzoai/ml/tree/main/hanzo-ml/src/vulkan) 04 / AppleMetalApple Silicon kernels and GPU-validated execution paths. [Explore Metal](https://github.com/hanzoai/ml/tree/main/hanzo-metal-kernels) ### Intelligence that connects to every layer. Agent capabilities sit alongside the inference path; not every request crosses every subsystem. Retrieve / remember #### Semantic + Graph Semantic retrieval, durable memory, provenance and connected knowledge for agent work. 2.0× multi-hop evidence recall vs cosineLoCoMo, held-out test split: every evidence turn of a multi-hop question in the top 20 for 37.5% of questions, against 18.8% for cosine search. [Product](https://hanzo.ai/products/memory)[Docs](https://docs.hanzo.ai/docs/openapi/graph)[Source](https://github.com/hanzoai/semantic)[The LoCoMo run](https://hanzo.ai/benchmarks/locomo) Understand / edit #### Code Intelligence Symbols, ASTs, references and repository-aware code navigation for coding agents. [Product](https://hanzo.ai/dev)[Docs](https://docs.hanzo.ai/docs/openapi/code)[Source](https://github.com/hanzoai/lsp) Build / ship #### Hanzo Forge Source hosting, builds, CI/CD and deployments connecting agents to shipped software. [Product](https://forge.hanzo.ai)[Docs](https://docs.hanzo.ai/docs/cli/platform)[Source](https://github.com/hanzoai/forge) Edge / ingress + egressRequest boundary #### In. Out. Under control. Admission, auth, routing, streaming and responses are boundary functions around the stack rather than another inference step. Outbound calls carry delegated credentials that stay in KMS. [Product](https://hanzo.ai/ingress)[Ingress](https://github.com/hanzoai/ingress)[Egress](https://github.com/hanzoai/egress) Telemetry / observabilityEnd to end #### Everything leaves a trace. Follow a request across models, decisions, tools and infrastructure to cost, quality, latency, usage and evaluation. [Product](https://hanzo.ai/o11y)[Docs](https://docs.hanzo.ai/docs/openapi/o11y)[Source](https://github.com/hanzoai/o11y) Research / recursive improvement ### Not just running. Learning. [Research](https://hanzo.ai/research)[Benchmarks](https://hanzo.ai/benchmarks)[Harness](https://github.com/hanzoai/benchmarks) - 01 / ObserveMeasureTraces · regressions - 02 / ProposeHypothesizeCandidate upgrades - 03 / IsolateExperimentSandboxed variants - 04 / ProveBenchmarkQuality · latency · cost - 05 / ControlGate + promoteReview · rollback A modular feedback architecture, not a claim of unattended production self-modification. Only candidates that meet the relevant evaluation and release gates are promoted. 12 native layersInspectable product, docs and source links for each architectural surface. [Open source](https://hanzo.ai/open-source) 4 GPU backendsCUDA, ROCm, Vulkan and Metal; backend-specific validation, never blanket speed claims. [Kernel DSL](https://github.com/hanzoai/ml/tree/main/hanzo-kernel) Measure the workPublished measurements are scoped to a named model, workload, hardware and baseline. [Benchmarks](https://hanzo.ai/benchmarks) GPU targets are parallel hardware backends, not sequential stages. Zen is generative reasoning; Kai provides decision intelligence. ZAP is a native RPC protocol for supported interconnect paths. Research and evaluation form a gated improvement loop. “Native” describes Hanzo’s architecture surfaces; not every implementation, weight, driver or operating system is owned or open sourced by Hanzo. Compute ## One runtime for every kind of work. The same plane, held for a millisecond or for a month. What changes is how much of a machine the work needs, not what you have to learn. - FunctionsScale to zero between calls. - ContainersYour image, deployed and addressable. - SandboxesIsolated shells for agents and untrusted code. - MachinesA whole machine, by the second. - GPUsAccelerators attached to any of them. - KubernetesThe substrate all of it runs on. ContainersDeploy an image and invoke it. Deployments, triggers, secrets, metrics and logs are all part of the same surface, so a function you ship is a function you can watch.api.hanzo.ai/v1/functions Machines Sandboxes Containers — Hanzo [GPUs](https://hanzo.ai/cloud/gpus)[Machines](https://hanzo.ai/machines)[Containers](https://hanzo.ai/cloud/containers)[Functions](https://hanzo.ai/functions)[Edge](https://hanzo.ai/edge)[Applications](https://hanzo.ai/cloud/applications)[Clusters](https://hanzo.ai/cloud/clusters)[Kubernetes](https://hanzo.ai/cloud/kubernetes) - gatewayReady - payments-apiReady - checkout-webBuilding - docs-siteReady Every request enters here, so the limiter does too. `api.hanzo.ai/v1/functions`[Open in the console](https://hanzo.ai/home) Example project. Every name in it is one you would give your own. [Compute API](https://docs.hanzo.ai/docs/openapi/compute)[Function API](https://docs.hanzo.ai/docs/openapi/function)[Sandbox API](https://docs.hanzo.ai/docs/openapi/sandbox)[Machines](https://hanzo.ai/machines) Data ## The data primitives AI applications use. Each primitive is the right shape for its job, and they share one project, one key and one bill. ### Datastore Columnar analytics over everything you collected. ### Object Storage Objects, signed URLs, lifecycle rules. ### Base An application backend: records, rules, realtime. ### Memory What an agent is allowed to remember between runs. [Base API](https://docs.hanzo.ai/docs/openapi/base)[Object storage API](https://docs.hanzo.ai/docs/openapi/s3)[KV API](https://docs.hanzo.ai/docs/openapi/kv)[Hanzo Base](https://hanzo.ai/base) Anywhere ## One control plane. Any infrastructure. The API, the console, the identities and the bill stay the same when the machines underneath change. ### Hanzo Cloud Capacity we operate and meter, provisioned by the call. ### Your cloud account Your credentials and your machines, under the same control plane. ### Your cluster Attach Kubernetes you already run and address it like any other. [Compute API](https://docs.hanzo.ai/docs/openapi/compute)[Node API](https://docs.hanzo.ai/docs/openapi/node)[Visor](https://hanzo.ai/visor) Security ## Security is part of the cloud, not a product on top of it. An agent gets an identity because it runs here, and the trail is written because every call goes through the same door. ### Identity One issuer for people, programs and agents, each request scoped to the org it acts for. ### Permissions Access follows the resource, not the caller. ### Secrets Held in KMS and handed only to the environment that needs them. ### Audit Actions written to a trail you can query, attributable to whoever took them. [IAM API](https://docs.hanzo.ai/docs/openapi/iam)[KMS API](https://docs.hanzo.ai/docs/openapi/kms)[Audit API](https://docs.hanzo.ai/docs/openapi/audit)[Security](https://hanzo.ai/security) One API ## One API. Every capability. The same key, organization, project, policy and usage ledger behind every path. The interface changes; the system does not. - api.hanzo.ai/v1/models - api.hanzo.ai/v1/agent - api.hanzo.ai/v1/function - api.hanzo.ai/v1/provisioning/sql - api.hanzo.ai/v1/provisioning/vector - api.hanzo.ai/v1/s3/buckets - api.hanzo.ai/v1/… Five operations, one key, no glue. Each is checked against the published OpenAPI document on every build. One org, one afternoon: a spaceship shop that did not exist this morning. Start with the models. `GET /v1/models`Returns the list of available models from the routing table. Models Play [Every capability](https://docs.hanzo.ai/docs/openapi)[The API](https://hanzo.ai/api) Ship ## Push code. Get a production system. An address, a certificate, secrets, metrics, logs and its share of the bill come back with it, because none of them is a separate purchase. ### Push A commit to the forge starts the build. ### Build The image is built in the cluster, never on a laptop. ### Release It rolls out with its secrets and its address, ready to roll back. ### Observe Logs, metrics and traces from its first request. [Deploy API](https://docs.hanzo.ai/docs/openapi/deploy)[Registry API](https://docs.hanzo.ai/docs/openapi/registry)[CLI docs](https://docs.hanzo.ai/docs/cli)[Registry](https://hanzo.ai/registry) Open ## Your cloud should not lock you in. The platform is open source, and the image we operate is the image you can pull: the same API and console on your cluster, in your network or on a laptop. $ curl -fsSL https://hanzo.sh | sh $ hanzo serve cloud $ docker run -p 8080:8080 ghcr.io/hanzoai/cloud [Quickstart](https://docs.hanzo.ai/docs/quickstart)[Open source](https://hanzo.ai/open-source)[Source](https://github.com/hanzoai/cloud) Pricing ## Pay for what runs. Nothing is billed for existing. One organization balance and one usage ledger, and every line traces back to a request. ### Models By usage ### Compute By execution ### Storage By consumption ### Managed services By usage [Pricing docs](https://docs.hanzo.ai/docs/pricing)[Billing API](https://docs.hanzo.ai/docs/openapi/billing)[Plans and rates](https://hanzo.ai/pricing) Catalog ## Explore the platform. Every plane, and what is in it. ### AI [Models](https://hanzo.ai/models) · [Agents](https://hanzo.ai/agents) · [Inference](https://hanzo.ai/engine) · [Embeddings](https://hanzo.ai/cloud/embeddings) · [Providers](https://hanzo.ai/cloud/providers) ### Compute [GPUs](https://hanzo.ai/cloud/gpus) · [Machines](https://hanzo.ai/machines) · [Containers](https://hanzo.ai/cloud/containers) · [Functions](https://hanzo.ai/functions) · [Edge](https://hanzo.ai/edge) ### Data [Datastore](https://hanzo.ai/datastore) · [Object Storage](https://hanzo.ai/s3) · [Base](https://hanzo.ai/base) · [Memory](https://hanzo.ai/cloud/memory) ### Dev [CLI](https://hanzo.ai/cli) · [Playground](https://hanzo.ai/playground) · [Desktop](https://hanzo.ai/desktop) · [API Keys](https://hanzo.ai/cloud/api-keys) ### Observe [Logs](https://hanzo.ai/cloud/logs) · [Metrics](https://hanzo.ai/metrics) · [Dashboards](https://hanzo.ai/dashboards) · [O11y](https://hanzo.ai/o11y) · [AI Metrics](https://hanzo.ai/cloud/ai-metrics) ### Security [IAM](https://hanzo.ai/iam) · [Authz](https://hanzo.ai/authz) · [KMS](https://hanzo.ai/kms) · [Secrets](https://hanzo.ai/cloud/secrets) · [Audit](https://hanzo.ai/cloud/audit) ### Network [Gateway](https://hanzo.ai/gateway) · [DNS](https://hanzo.ai/dns) · [Service Mesh](https://hanzo.ai/cloud/service-mesh) · [Nodes](https://hanzo.ai/cloud/nodes) [The complete catalog](https://hanzo.ai/products) ## Build on the stack.Or run it yourself. Models, agents, compute, data, security and observability behind one API, from prototype to production. [Start building](https://platform.hanzo.ai/api-keys)[Read the docs](https://docs.hanzo.ai/docs/quickstart)