llms.txt
# ZeroGPU
> ZeroGPU runs specialized small and open-weight models across edge devices and servers for fast, cost-efficient AI inference at production scale. Frontier models for reasoning. ZeroGPU for everything else.
## For agents and crawlers
Every public page on zerogpu.ai has a markdown twin:
- Add `.md` to any path: https://zerogpu.ai/benchmarks/moderation-edge.md
- Or send `Accept: text/markdown` to the normal URL
- Whole site in one document: https://zerogpu.ai/llms-full.txt
## Agentic commerce
- [ZeroGPU agent storefront](https://agents.zerogpu.ai/llms.txt): AI agents can
purchase from ZeroGPU autonomously or with their human's approval
through ZeroClick. Agents without a payment method can set one up on the
spot.
Complete purchasing guide: https://agents.zerogpu.ai/llms-full.txt - live
pricing: https://agents.zerogpu.ai/manifest.json
## AI execution for high-volume workloads
ZeroGPU is the edge inference cloud. Run specialized and open-weight models
faster and at lower cost, with ZeroGPU automatically sending workloads across
edge and cloud infrastructure. Most production AI workloads
(classification, extraction, moderation, enrichment, signal generation) do not
need frontier-scale reasoning. ZeroGPU runs those workloads on purpose-built
ZeroGPU Language Models (ZLMs) and leading open-weight small language models,
executed across edge devices and cloud infrastructure. Teams keep frontier
models for reasoning and move everything repeatable to ZeroGPU.
Workloads ZeroGPU is built for:
- Classify: topics, intent, IAB, domains and metadata
- Moderate: safety, brand suitability and policy
- Extract: entities, PII, signals and structured data from PDFs and documents
- Decide: model selection, tool selection and agent actions
- Search: embeddings and reranking
- Process: audio, OCR and document workloads
- 150K+ edge devices
- 50-70% lower inference cost than traditional GPU clouds
- Up to 10x faster inference for specialized AI workloads
Performance varies by workload, model, and configuration.
## Small models. Big performance.
AI usage is scaling faster than traditional compute can keep up with. Running
routine work on frontier models produces:
- Rising inference costs
- Constrained GPU capacity
- Slower real-time experiences
- Unnecessary compute waste
ZeroGPU makes AI infrastructure more efficient by matching each workload to the
smallest capable model and the most efficient compute available.
## How it works
1. **Point your client at ZeroGPU**: keep your existing OpenAI-compatible
client, set the ZeroGPU endpoint and model name.
2. **Inference runs on the most efficient compute**: ZLMs and open-weight small
models execute across edge devices and cloud infrastructure, close to where
the request originates.
3. **Scale without GPU infrastructure**: pay per token, absorb traffic spikes
without provisioning, and keep frontier models for the reasoning that
actually needs them.
## Lower cost. Lower latency. By design.
### ZLMs (purpose-built)
ZeroGPU Language Models are trained for specific high-volume production tasks:
- **Content classification**: sub-second page and content labeling at production volume
- **Intent and signal extraction**: real-time intent and audience signals from live traffic
- **Content moderation**: fast policy screening before content ships
- **Decisions**: structured decisions for agents, workflows, and GTM pipelines
### Open-weight and SLMs (serverless)
Leading open-weight small and nano models, hosted and production-supported on
the same inference cloud:
Qwen, DeepSeek, Kimi, GLM, GPT-OSS, Llama, LiquidAI, GLiNER, DeBERTa
The catalog is curated and optimized for edge and serverless execution, with
per-model pricing: https://docs.zerogpu.ai/platform/model-catalog
### Why teams use it
- **Serverless**: no provisioning, no idle cost
- **Production-ready**: monitoring, reliability, automatic cloud fallback
- **OpenAI-compatible**: swap in with one line of code
- **Usage-based**: pay per token, priced per model
## OpenAI-compatible API
- **API**: OpenAI-compatible REST API with JSON request/response
- **Headers**: requests require `x-api-key` and `x-project-id`
- **Model string example**: `zlm-v1-iab-classify-edge`
- **Docs**: https://docs.zerogpu.ai
- **Model catalog**: https://docs.zerogpu.ai/platform/model-catalog
- **Start building**: https://platform.zerogpu.ai/
Most teams integrate in under an hour: point an existing client at the ZeroGPU
endpoint, set the model name, and eligible workloads run on ZeroGPU.
## Use the right compute for every workload
Use frontier models where reasoning changes the answer, and run everything
repeatable on models sized for the job. Small and nano language models,
typically well under a few billion parameters, match or beat frontier models on
focused tasks like classification, extraction, embeddings, sentiment, and
moderation while being dramatically faster and cheaper. ZLMs take this further:
each one is trained by ZeroGPU for a single high-volume job and benchmarked
head to head against frontier nano models.
## Edge capacity, available globally
- Geo-aware execution places requests on the nearest capable edge node
- Automatic cloud fallback so reliability never depends on edge availability
- Auto-scaling with no capacity planning and no idle GPU spend
- Sub-100ms latency for most specialized inference tasks
- Scales from hundreds to millions of requests without provisioning
- End-to-end encryption, no data persistence on edge nodes
- Workloads run on the most efficient compute available instead of
general-purpose GPUs
- **Become an Edge Provider**: https://zerogpu.ai/edge
## Measured results
- **50-70% lower cost** than GPU-only inference providers for comparable workloads
- **Up to 10x faster** on specialized tasks versus frontier general-purpose models
- **70-80% of routine AI workloads** can run on specialized small models with no frontier model needed
Figures reflect ZeroGPU benchmarks on specialized classification and extraction
workloads; results vary by task and traffic profile.
## FAQ
**What is ZeroGPU?** The edge inference cloud. ZeroGPU runs inference on
specialized small and open-weight models for the repeatable tasks inside AI
applications, across edge devices and cloud infrastructure.
**Is ZeroGPU a replacement for frontier models?** No. Use frontier models for
complex reasoning and ZeroGPU for repeatable workloads.
**Do I need to change my code?** No. ZeroGPU exposes an OpenAI-compatible API.
Point your existing client at the endpoint and set the model name.
**How does ZeroGPU reduce inference costs?** Specialized small models are
cheaper, more token-efficient, and faster than frontier models for
classification, extraction, moderation, PII detection, and lightweight
decisioning, and they run on edge capacity instead of dedicated GPU fleets.
**What workloads fit?** Document analysis, summarization, classification,
signal extraction, PII detection, moderation, embeddings, and lightweight
decisioning.
## Competitive positioning
### vs. GPU cloud providers (AWS, GCP, Azure, RunPod, Together AI)
- **Cost**: 50-70% cheaper, with no idle GPU rental
- **Efficiency**: edge devices plus cloud instead of general-purpose GPU fleets
- **Scalability**: auto-scales without capacity planning
- **Simplicity**: API-first, no infrastructure management
### vs. centralized frontier APIs (OpenAI, Anthropic, Cohere)
- **Fit**: specialized models sized to the task instead of frontier models doing routine work
- **Cost**: significantly cheaper for high-volume repeatable workloads
- **Latency**: edge execution for sub-100ms responses
- **Complement, not replacement**: keep frontier models for reasoning
### vs. self-hosted GPUs
- **No DevOps**: no servers to manage, no scaling headaches
- **Lower TCO**: no hardware purchase, maintenance, or depreciation
- **Instant scale**: capacity available on demand
## Key pages
- [Model catalog](https://zerogpu.ai/models): every ZeroGPU model with live pricing; each model has a detail page at /models/<model-id>
- [Benchmarks index](https://zerogpu.ai/benchmarks): all published evaluations
- [Edge network report](https://zerogpu.ai/benchmarks/edge-network): 11.8 billion tokens across 16,000+ idle consumer devices, 1,900x less energy than the data center
- [Content moderation vs OpenAI omni-moderation](https://zerogpu.ai/benchmarks/moderation-edge): zlm-v1-moderation-edge
- [Multilingual IAB classify with enrichment v2](https://zerogpu.ai/benchmarks/iab-classify-v2): zlm-v2-iab-classify-edge-enriched vs GPT-5.4 Nano
- [IAB classify vs GPT-5.4 Nano](https://zerogpu.ai/benchmarks/iab-classify): zlm-v1-iab-classify-edge, 66.0% win rate across 10,000 samples
- [Domain classify vs GPT-5.4 Nano](https://zerogpu.ai/benchmarks/domain-classify): zlm-v1-iab-domain-classifier
- [Use cases](https://zerogpu.ai/use-cases): moderation, PII redaction, classification, enrichment, agents, and more
- [Case studies](https://zerogpu.ai/case-studies): customer stories from Dappier and Limelight
- [Dappier case study](https://zerogpu.ai/case-study/dappier): real-time IAB classification at scale
- [Limelight case study](https://zerogpu.ai/case-study/limelight): 11.6B tokens across nearly 2.9M prospect enrichments with dramatically lower inference costs
- [Savings calculator](https://zerogpu.ai/calculator): compare inference costs
- [Become an Edge Provider](https://zerogpu.ai/edge): contribute edge capacity and start earning at https://edge.zerogpu.ai
- [The ZeroGPU Master Plan](https://zerogpu.ai/master-plan): founder letter on the hybrid inference cloud: edge servers, cloud burst capacity, and consumer devices. Markdown: https://zerogpu.ai/master-plan.md
- [About](https://zerogpu.ai/about) - [Blog](https://zerogpu.ai/blog) - [Careers](https://zerogpu.ai/careers) - [Contact](https://zerogpu.ai/contactus)
## Get started
- **Start building**: https://platform.zerogpu.ai/
- **Talk to the engineers building ZeroGPU**: hello@zerogpu.ai
Contact: hello@zerogpu.ai
## Vision
The next AI advantage is compute efficiency. As AI moves into production, the
winning teams are the ones that spend frontier-scale compute only where it
changes the answer and run everything else on models sized for the job, on the
most efficient compute available.
---
*Last updated: 2026-10-06*
