About ZeroGPU
Overview
ZeroGPU runs repeatable, high-volume AI workloads across a hybrid network of edge devices and cloud infrastructure. It serves purpose-built ZeroGPU Language Models (ZLMs) trained for specific production tasks such as content classification, intent and signal extraction, moderation and structured decisions, alongside popular open-weight small and nano models hosted serverless on the same platform.
An OpenAI-compatible API lets teams send selected workloads to specialized models with one line of code. For each request, ZeroGPU matches the workload to the right model and chooses the execution location based on capability, latency, availability and cost, then falls back to cloud capacity when edge execution is not appropriate. The network spans more than 150,000 edge devices and regions from San Francisco and New York to London, Dubai, Mumbai, Singapore, Tokyo and Sydney.
Key Benefits
- Cut inference cost by 50–70% versus traditional GPU clouds on repeatable workloads
- Run specialized AI workloads up to 10x faster through edge execution
- Reach 150K+ edge devices spread across North America, South America, Europe, the Middle East, Asia and Australia
- Swap in an OpenAI-compatible endpoint without rebuilding the existing application
- Avoid provisioning work and idle cost with serverless model hosting
- Keep workloads running through cloud fallback when edge execution is not appropriate
How It Works
Teams browse the model catalog, then send requests to the OpenAI-compatible endpoint at api.zerogpu.ai/v1/chat/completions with an API key and project ID. ZeroGPU selects the model that fits the workload and routes it to the best execution location across edge devices or cloud, with streaming, structured output and batch support available. Agents can also buy capacity directly from the agent storefront, which publishes prices and payment instructions and handles payment itself, so an agent can purchase autonomously or with its human's approval and without signup.
Use Cases
- Content and marketing teams label pages, topics, domains, intent and IAB categories at production volume
- Trust and safety teams screen content for safety, brand suitability and policy before it ships
- Document teams extract entities, PII and structured data from PDFs and run OCR and audio workloads
- Agent developers get structured decisions for model selection, tool selection and agent actions
- Search and GTM teams run embeddings and reranking and pull real-time intent and audience signals from live traffic
- Device owners and app teams contribute idle compute through the Edge SDK and share in the inference revenue their devices generate
Why Choose This Product
ZeroGPU pairs a purpose-built model line with a distributed inference delivery network, so repeatable tasks like classification and moderation do not carry frontier-model economics. The OpenAI-compatible API keeps existing SDK workflows and request patterns intact, usage-based per-token pricing scales with actual volume, and cloud fallback covers the cases where edge capacity is unavailable or unsuitable.
ZeroGPU Pros & Cons
- Distributed network of 150K+ edge devices across 11 listed regions
- States 50–70% lower inference cost than traditional GPU clouds
- OpenAI-compatible API with one-line code swap and existing SDK support
- Cloud fallback when edge execution is unavailable or unsuitable
- Per-token, per-model pricing with no provisioning or idle cost
- Performance varies by workload, model and configuration
Key Features
Purpose-Built ZLMs
ZeroGPU Language Models are trained for specific high-volume production tasks including content classification, intent extraction, moderation and structured decisions for agents and workflows.
Open-Weight Hosting
The platform hosts popular open-weight small and nano models such as Qwen, DeepSeek, Kimi, GLM, GPT-OSS, Llama, LiquidAI, GLiNER and DeBERTa on a serverless, production-supported service.
OpenAI-Compatible API
Requests go to the api.zerogpu.ai/v1/chat/completions endpoint using familiar API patterns, existing SDK compatibility, streaming, structured output and batch support.
Hybrid Routing
ZeroGPU picks the execution location for each workload across 150K+ edge devices and cloud based on capability, latency, availability and cost.
Cloud Fallback
Cloud execution keeps workloads running when edge capacity is unavailable or when edge execution is not appropriate for the request.
Usage-Based Pricing
Customers pay per token with prices set per model rather than per provisioned GPU.
Edge SDK
Device owners and app teams add the Edge SDK to turn available device capacity into part of the global inference network or orchestrate a private fleet.
Agent Storefront
The storefront at agents.zerogpu.ai publishes prices, payment instructions and setup so AI agents can purchase capacity autonomously or with human approval and no signup.
Frequently asked questions about ZeroGPU
What is ZeroGPU?
ZeroGPU describes itself as an edge inference cloud for AI execution on high-volume workloads. It runs purpose-built ZeroGPU Language Models and popular open-weight models across a hybrid network of edge devices and cloud infrastructure, and it exposes the service through an OpenAI-compatible API.
What workloads should I run on ZeroGPU?
ZeroGPU targets repeatable AI work you run millions of times: classifying topics, intent, IAB categories, domains and metadata; moderating for safety, brand suitability and policy; extracting entities, PII and structured data from PDFs and documents; producing decisions for agents and workflows; embeddings and reranking; and audio, OCR and document processing.
What models are available on ZeroGPU?
The platform serves purpose-built ZLMs for tasks such as content classification, intent extraction, moderation and decisions, plus open-weight small and nano models including Qwen, DeepSeek, Kimi, GLM, GPT-OSS, Llama, LiquidAI, GLiNER and DeBERTa. The model catalog lists what is available and priced per model.
Is ZeroGPU OpenAI-compatible?
Yes. ZeroGPU exposes an OpenAI-compatible API, so teams send selected workloads to specialized models using familiar API patterns and existing SDKs without rebuilding the application. The API supports streaming, structured output and batch requests.
Do I need to manage edge infrastructure?
No. ZeroGPU offers serverless hosting with no provisioning and no idle cost, and it handles production monitoring, reliability and cloud fallback so customers run workloads without operating edge hardware themselves.
What happens when edge capacity is unavailable?
ZeroGPU dynamically routes eligible workloads across edge devices and cloud infrastructure. Cloud fallback keeps workloads running when edge capacity is unavailable or when edge execution is not appropriate for a given request.