Sign in to Hivenet
Access your files and manage your storage account.
Store
Access your files and manage your storage account.
Compute
Launch instances and manage your compute resources.
Store and Compute still use separate accounts today. We’re working on a unified sign-in.
Just need to send a file?
Use Send directlyNew to Hivenet? Get Started
Get started with Hivenet
Choose what you want to use first.
Store
Back up photos and files across your devices.
Compute
Launch self-serve GPU compute for demanding work.
Store and Compute still use separate accounts today. We’re working on a unified account experience.
Need Hivenet for business?
Talk to salesNeed to send a file?
Open SendAlready have an account? Sign in
Compute - Inference API
Run open-weight models through dedicated, managed endpoints.
Deploy production inference without operating the serving stack yourself. Hivenet Inference API gives you an OpenAI-compatible endpoint, dedicated replica capacity, and regional deployment on Hivenet infrastructure.
OpenAI-compatible API
Dedicated endpoints
1-4 replicas
Per-replica pricing
Billed by the second
Regional deployment
Qwen · Llama · Mistral · Falcon · GPT OSS · Gemma
Managed inference with dedicated capacity and a familiar API.
Running inference yourself gives you control, but it also leaves your team operating GPUs, serving engines, routing, authentication, monitoring, and capacity. Managed model APIs remove most of that work, but often give you less control over infrastructure placement and how you pay for sustained workloads.
Hivenet Inference API sits between those two approaches.
Hivenet operates the serving layer.
Choose a model and the capacity you need. Hivenet runs the infrastructure, router, serving environment, endpoint, and operational layer behind it.
The API experience stays familiar.
Use an OpenAI-compatible endpoint with common client patterns. Point your application at the Hivenet base URL, authenticate with an API key, and call the model.
The cost follows dedicated capacity.
Pay for the replica capacity behind the endpoint rather than a separate meter for every input and output token. Runtime is billed by the second.
For teams running production inference at real volume.
Hivenet Inference API is for teams that want to use open-weight models in production without becoming responsible for the infrastructure that serves them.
Production teams with steady inference traffic
Use dedicated capacity for workloads where request volume is sustained enough that reserving the endpoint makes sense.
SMBs with growing AI bills
Move suitable production workloads from per-token APIs to infrastructure-backed pricing you can see before you deploy.
Teams with residency needs
Choose from available deployment locations and keep the endpoint tied to the selected infrastructure location.
Developers using OpenAI-compatible tooling
Keep familiar request patterns with the OpenAI SDK and other compatible clients instead of rebuilding the application around another proprietary API.
Managed endpoint or compute instance?
Use Hivenet Inference API when you want the endpoint. Use Compute with Hivenet when you want the machine and full control over the runtime. Use Hivenet Router when you want to operate the routing layer yourself.
Need
Use
Why
I want an OpenAI-compatible endpoint
Hivenet Inference API
Hivenet operates the serving layer and endpoint
I want predictable dedicated inference capacity
Hivenet Inference API
Choose replica capacity and pay for its runtime
I want to run vLLM, SGLang, llama.cpp, PyTorch, custom models, or fine-tuning myself
GPU/CPU rental with Hivenet
You control the instance, runtime, and software stack
I want to route across inference engines I operate myself
Hivenet Router
You run the open-source routing layer and control its policies
I need a custom AI system around sensitive data or unusual deployment requirements
Private AI
Hivenet can help scope the model, data, infrastructure, and support path
Swipe left to see more
Built for the AI workloads teams actually run.
Open-weight models can handle a wide range of production tasks when the model is matched to the workload and tested against real data.
RAG
Serve retrieval-augmented generation workflows for internal knowledge, customer support, documentation, and business data.
Structured extraction
Extract dates, entities, categories, and structured fields from documents, messages, tickets, invoices, or records.
Summarization
Summarize documents, conversations, support threads, research, and operational content.
Classification
Classify messages, records, tickets, documents, and workflow inputs through a dedicated endpoint.
Code assistance
Run code-related workflows where the available model meets your quality, context, latency, and cost requirements.
Internal tools
Build assistants, automation, and internal AI features without operating the inference stack behind them.
From model to endpoint in a few steps.
Inference API is available from the Hivenet Compute console. Choose the deployment you need; Hivenet handles the serving environment behind it.
Choose a location
Select from the deployment locations currently available for your workload.
Choose a model and variant
Pick a model from the managed catalog, then choose the available variant that fits your context, quality, performance, and cost requirements.
The variant determines the underlying serving configuration. You do not need to configure the GPU, precision, or serving engine yourself.
Set replica capacity
Choose between 1 and 4 replicas, subject to available capacity.
The console shows the expected throughput and price as you adjust the deployment.
Review and deploy
Name the endpoint and review the model, location, replicas, hourly rate, and projected cost before starting it.
Create a key and connect
Generate an API key for your organization and use the deployment URL with an OpenAI-compatible client.
Dedicated endpoint controls you can see.
The endpoint exposes the important operational choices instead of hiding them behind a generic API plan.
Capacity
1–4 replicas
Set the replica count when you deploy and change it later without rebuilding the endpoint.
Billing
Per second
Replica runtime is metered by the second rather than rounded into a monthly commitment.
Integration
OpenAI-compatible
Use familiar OpenAI-compatible clients and request patterns with the Hivenet endpoint.
Routing
One router per customer
Your deployments share a customer-specific router rather than a single router shared with unrelated customers.
Lifecycle
Start · stop · terminate
Stop capacity when you do not need it, start it again later, or terminate the deployment when the workload is finished.
A curated model catalog for production workloads.
Hivenet Inference API starts from a managed catalog rather than an empty GPU machine. Exact models, variants, and locations depend on current availability and benchmark readiness.
Model family
Good starting point for
What to evaluate
Qwen
Extraction, RAG, structured output, multilingual workloads
Context, quality, throughput, structured output
Llama
RAG, summarization, assistants, internal tools
Context, latency, tool support, workload fit
Mistral
Instruction workloads, summarization, tools
Context, throughput, tool support
Falcon
Efficient general-purpose inference
Model size, throughput, workload fit
GPT OSS
General instruction and reasoning workloads
Quality, model size, latency
Gemma
Compact model workloads and experimentation
Model size, quality, throughput
Swipe left to see more
Qwen for production extraction and RAG
Qwen is a strong starting point for teams testing structured extraction, RAG, and production workflow automation.
Llama and Mistral workloads
Run widely adopted model families for RAG, summarization, internal tools, and model-serving experiments.
Need your own weights?
The managed catalog is the starting point for Inference API.
If you need custom weights, fine-tuned models, or direct control over the serving environment, use Compute with Hivenet or talk to Hivenet about a custom deployment.
Predictable pricing for dedicated endpoints.
Hivenet Inference API uses per-replica pricing billed by the second. You pay for provisioned inference capacity while it is running, rather than paying separately for each token generated.
Per-replica pricing
The model variant determines the infrastructure configuration and rate behind each replica.
Billed by the second
Pay for actual replica runtime. Stop the deployment when you no longer need the capacity and billing pauses once it has stopped.
Tokens included
Input and output tokens are included in the dedicated endpoint rate for the current product.
API calls and egress included
There is no separate charge for API calls or egress in the current dedicated endpoint pricing model.
Tier
Example use
Price
1 × RTX 5090
Smaller models and variants that fit a single GPU
from €1.10/hr
2 × RTX 5090
Production variants needing more memory or throughput
from €2.10/hr
4 × RTX 5090
Larger or higher-capacity model variants
from €3.80/hr
Swipe left to see more
Regional infrastructure for dedicated AI endpoints.
Hivenet Inference API runs on Hivenet-operated, Policloud-backed infrastructure. Deployment location remains part of the product rather than disappearing behind an abstract API endpoint.
One location in the current release
Choose an available location for your first Inference deployment. Subsequent deployments currently inherit that location.
Single-tenant routing
One router per customer keeps the routing path isolated from unrelated customer traffic and easier to understand operationally.
Full-stack operation
Hivenet operates the router, serving runtime, endpoint infrastructure, authentication layer, monitoring, and billing behind the managed service.
We own the hardware
Run inference on a Policloud-backed path instead of routing production AI entirely through default hyperscaler APIs.
Test quality and throughput on your real workload.
Inference performance depends on the model, variant, context, request shape, output length, replica count, and traffic pattern. A benchmark is useful only when those conditions are visible.
Benchmarks with context
See what was actually tested
Hivenet benchmark results should show the model, variant, hardware configuration, request rate, and workload conditions behind the result.
Model evaluation
Your prompts decide the fit
Test the model against your real inputs, expected outputs, quality requirements, latency target, and traffic before moving production volume.
Endpoint metrics
Watch the signals that matter
Track requests, tokens, latency, errors, cost, and time to first token where available.
Keep your client code familiar.
Hivenet Inference API uses an OpenAI-compatible API so existing integrations can usually keep the same client pattern and change the endpoint configuration.
Point the client at your deployment
Each deployment has its own endpoint URL. Use it as the base_url for the client.
Use an organization API key
Create a Hivenet Inference API key and use it to authenticate calls to your organization's endpoints.
Send the model actually deployed
Specify the served model name in the request rather than relying on a generic default.
# Use the OpenAI client, pointed at Hivenet
from openai import OpenAI
client = OpenAI(
api_key="HIVENET_API_KEY",
base_url="https://api.hivenet.example/v1"
)
response = client.chat.completions.create(
model="qwen-example",
messages=[{"role": "user", "content": "Summarize this document."}]
)
Built around real production needs.
Hivenet Inference API is designed for production workloads where dedicated capacity, model choice, cost visibility, and infrastructure placement matter.
Production workload
High-volume document automation
A business automation customer uses a dedicated Qwen endpoint for part of a production extraction workflow.
The useful question is not whether an open model can replace every API call. It is whether a tested model can handle a specific production workload well enough to move that traffic onto dedicated capacity.
Best-fit buyer
Teams with meaningful, steady API spend
Dedicated endpoints make the most sense when a team already has production traffic and can identify workloads with sufficiently predictable demand.
Variable experiments and small sporadic workloads may be better suited to a different infrastructure path.
Need a different AI infrastructure path?
Hivenet Inference API is the managed endpoint path. If you want to operate more of the stack yourself, or need a different data and infrastructure setup, use the Hivenet product that matches the job.
GPU/CPU rental
Rent RTX 5090 or vCPU instances when your team wants full control over the instance, framework, and serving stack.
Private AI
Work with Hivenet on guided AI projects involving sensitive data, model choice, deployment support, or custom requirements.
RAG
Build retrieval systems on your own data using Hivenet's AI and storage paths.
S3-compatible storage
Store datasets, documents, and AI pipeline artifacts with S3-compatible tools and free egress.
FAQ
Common questions
Move one production AI workload to Hivenet.
Bring your current API usage, model needs, latency target, region requirements, and quality bar. We'll help you decide whether a managed foundational model endpoint is the right path.
