We run the infrastructure,

you ship the models.

— The faster way to train, ship and scale your AI.

Arc-7 Inference

// Latest release

See Arclight in action

A guided tour of the console, live, with a model of your own.

Book a demo

99.98%

Uptime

90-day average

+38.4%

Throughput

after auto-tuning

Running in production at

▲ Northwind

● lumen health

KESTREL

◯ Orbital

paperplane

Halcyon

✦ QUILL&CO

Driftwood

Our services

002

arc—light

Everything between a trained model and a product people use.

Four services, one console, one bill.

Inference

Serverless endpoints that answer in milliseconds and scale to zero when nobody is asking.

inference

Training

Fine-tune on reserved H200 clusters, checkpointed every ten minutes, never lost to a bad node.

training

Guardrails

Every prompt and every answer filtered, logged and kept in the region you choose.

protection

Scale

Autoscaling across four regions, billed by the second, from the first user to the millionth.

growth

From key to production

003

arc—light

If your code talks to a model today, it talks to Arclight this afternoon.

01

Pick a model

Forty models ready to call, or your own weights uploaded from Hugging Face or a bucket.

02

Change one line

Arclight speaks the OpenAI format: point your client at our address and keep your code.

03

Ship, then scale

Traffic grows, GPUs follow, in four regions. You see every call, its cost and its trace.

Read the quickstart

app.py
from openai import OpenAI

client = OpenAI(
    base_url="https://api.arclight.ai/v1",
    api_key="ARCLIGHT_API_KEY",
)

answer = client.chat.completions.create(
    model="arc-7",
    messages=[{"role": "user", "content": "Summarise this contract"}],
)
print(answer.choices[0].message.content)

An agent, at work

004

arc—light

Agents on Arclight plan, call your tools and answer, and every step is kept as a trace you can replay.

Try the console

You

Why did checkout errors rise on Tuesday?

01

Plan

Read the error logs, compare deploys, check the payment provider.

02

logs.search

status:500 route:/checkout since:tue → 1,284 errors

03

deploys.list

web@4.18.2 shipped Tue 09:12, 3 minutes before the first error

04

code.diff

currency rounding moved to the client in cart.ts

Arc-7

The errors began with web@4.18.2: carts in yen were rounded twice, and the payment provider refused the totals. Reverting cart.ts line 88 clears it.

Model

Arc-7

Steps

4

Time

6.2 s

Cost

$0.011

4

Regions

Virginia, Oregon, Amsterdam, Singapore

38K

GPUs online

H200, H100 and L40S

2.1T

Tokens a day

served for 3,000 teams

180ms

To the first token

Arc-7, at the 50th percentile

Models, ready to call

005

arc—light

Ours and the best open weights, served on the same engine and billed the same way.

All models

“We moved four models to Arclight in a week and cut our inference bill by 61%. The traces alone paid for it: our agents finally make sense.”

Maya Lindqvist

Head of AI, Northwind

Read the story

Start free, grow on usage

006

arc—light

Every plan runs every model. You pay for what you call, and for the GPUs you keep.

Compare plans

Build

$0

plus usage

For a side project or a first prototype.

  • $10 of credit to start
  • Every model, serverless
  • Pay per token or per second
  • Community support
Start free

ScaleMost teams

$499

per month, plus usage

For a team with a product in production.

  • Reserved GPUs at 30% off
  • Fine-tuning and private models
  • Guardrails and audit logs
  • 99.95% uptime, in writing
  • Support in one business hour
Start a trial

Enterprise

Custom

billed yearly

For a company with its own rules.

  • Dedicated clusters, any region
  • Your cloud or ours (VPC)
  • SSO, SCIM and data residency
  • 99.99% uptime and a named engineer
Talk to sales

Create a free website with Framer, the website builder loved by startups, designers and agencies.