v4 Cold starts under 40 ms in 31 regions · changelog

Run your workload where
your users already are

Deploy any workload to the edge in one command. No clusters, no warm pools, no capacity planning — you pay for what you serve.

$npx argosy up --model llama-3.3-8b

Serving 2.41B requests a month for engineering teams at

Northwind Ravelin Kestrel Health Bellwether Ostara Ferrix
The network

Latency is a distance problem before it is a compute problem

A workload answers in about 30 ms when it runs next to the user. Send that same request to a datacentre on another continent and back, and the round trip alone costs three times that. So we put the work next to the user instead.

01

Workloads stream, not deploy

Workloads are content-addressed and pulled on first request. A region that has never served yours is warm in under 400 ms and cached thereafter.

02

Route by residency, not by round-robin

Pin a workload to the EU, to a single country, or to hardware you own. The router honours it before it honours latency.

03

Scale to zero, honestly

No idle compute-hour billing and no warm-pool minimum. If nobody called your workload last month, the invoice says nothing.

The API

One endpoint. The SDK you already import.

Argosy speaks the OpenAI wire format, so migration is a base URL and a key. Streaming, tool calls, structured output and vision all behave the way your client library expects.

import OpenAI from "openai";

const argosy = new OpenAI({
  baseURL: "https://edge.argosy.dev/v1",
  apiKey:  process.env.ARGOSY_KEY,
});

// routed to the nearest warm region
const stream = await argosy.chat.completions.create({
  model:  "llama-3.3-8b-instruct",
  stream: true,
  messages: [{ role: "user",
    content: "Summarise this ticket in one line." }],
});
Responseiad1 · 31ms to first token
data: Customer cannot complete checkout when a saved card has expired ; retry loop never surfaces the error to the UI .
data: [DONE]
ttft 31mstokens 28 tok/s 142cost $0.000041
Request tracereq_9c4e17 · 41ms total
Edge accept2.4ms
Auth & quota1.6ms
Router decision1.1ms
KV cache lookup3.8ms
Prefill9.7ms
Decode · 28 tok19.8ms
Egress2.6ms
region iad1cache hit queue 0msretries 0
Observability

Every request is a trace, not a line in a log file

You cannot tune what you cannot see. Argosy emits a span for each stage of the request — including the ones most platforms hide, like queue time and cache behaviour.

Queue time is reported separately

If your p99 is bad because you are rate-limited rather than because the workload itself is slow, the trace says so. No inferring it from wall clock.

Traces are OTLP

Ship them to whatever you already run. There is no proprietary agent and no second dashboard you are obliged to look at.

Getting there

Three commands, then it is somebody else’s problem

01

Point it at a model

Any open-weights checkpoint from the registry, or your own from S3.

$ argosy up --model llama-3.3-8b
resolving weights … ok
quantising fp8 … ok
dep_7fa21c created
02

Choose where it may run

Regions, residency rules and spend ceilings live in one file.

regions: ["eu-*", "iad1"]
residency: "eu-only"
max_spend: 2000
scale_to_zero: true
03

Send it traffic

Swap the base URL. Nothing else in your application changes.

$ curl edge.argosy.dev/v1/chat \
  -H "authorization: Bearer …"
200 OK · iad1 · 31ms
x-argosy-cache: hit
38ms

p50, region with nothing cached yet

31

6 with dedicated, reservable capacity

99.99%

Measured at the edge, not the origin

4.2×

Than always-on reserved capacity at our median load

Pricing

You pay for tokens. That is the whole model.

No per-seat fee, no reserved instances you forget to turn off, and no charge for a region simply existing.

Hobby

$0

1M tokens a month, then usage rates. For side projects and evaluation.

  • All 31 regions
  • Every open-weights model
  • Community support
  • Scale to zero
Start free

Team

$0.11/ M tokens

Metered per token served, billed to the second. Most teams land here.

  • Residency & routing rules
  • OTLP trace export
  • Spend ceilings and alerts
  • 99.9% SLA
Start free

Dedicated

Talk to us

Reserved GPUs in named regions, or your own hardware under our control plane.

  • Reserved capacity
  • Private networking
  • 99.99% SLA
  • Named support engineer
Contact sales

Put the model next to the user.

One command, thirty-one regions, and an invoice that matches what you actually served.