002
Platform
— Everything between a model and its users.
Inference, training, guardrails and scale, built as one platform so a model you fine-tune this morning serves traffic this afternoon.
Four services, one console
002
arc—light
Each one stands alone. Together they take a model from a notebook to a product with millions of users.
01 Inference
Endpoints that answer before the user notices
Every model runs on our own inference engine: speculative decoding, paged attention and a prompt cache that keeps your system prompt warm between calls. Endpoints scale to zero when nobody is asking and come back in under a second.
180 ms to the first token on Arc-7
Serverless, or reserved GPUs at 30% off
Batch API at half the price per token
OpenAI-compatible, so one line moves you
02 Training
Fine-tuning that survives a bad node
Upload a dataset, choose a base model and a budget, and Arclight runs the job on reserved H200 clusters. A checkpoint is written every ten minutes, so a run that loses a node resumes on its own and you pay only for the minutes that counted.
LoRA, full fine-tuning and DPO
Evaluations after every epoch
Automatic resume, no charge for lost minutes
Your weights stay yours, exportable any time
03 Guardrails
Policies written in plain words
Write what a model must never say in a sentence. Guardrails turns it into a check that runs on every prompt and every answer in under 10 ms, masks personal data before the model sees it and logs every block with its reason.
Policies in plain English
PII masked before inference
Every call kept in the region you choose
Audit logs exported to your SIEM
04 Scale
Four regions, billed by the second
Traffic is routed to the nearest region with room to spare, and capacity follows demand across Virginia, Oregon, Amsterdam and Singapore. You see every call, its cost and its trace in one console.
Autoscaling with no cold starts on reserved GPUs
Failover between regions in seconds
Spend limits per key, per team, per project
99.99% uptime on Enterprise, in writing
How a call travels
003
arc—light
Every request passes one gateway, one router and lands on the pool nearest your user with room to spare.
Median path: 4 ms at the gateway, 1 ms in the router, the rest is the model.
Secure by default
004
arc—light
Encrypted in transit and at rest, isolated per customer, and never used to train anything.
SOC 2 Type II
Audited every year by an independent firm; the report is yours under NDA.
ISO 27001
An information security system certified across every region we run.
GDPR EU data
A data processing agreement, EU residency in Amsterdam, no training on your data.
HIPAA Ready
A business associate agreement for teams who handle health records.
Plugs into your stack
005
arc—light
The tools your team already runs, connected in the console or with one environment variable.
Hugging Face
Models
LangChain
Agents
LlamaIndex
Retrieval
Vercel
Deploy
Datadog
Observability
Snowflake
Data
Pinecone
Vectors
Weights & Biases
Training
GitHub Actions
CI
Terraform
Infrastructure
Okta
Identity
Slack
Alerts