Fictional demo — AXION Corp. does not exist. Every product, price, benchmark and service on this site is invented test data.

AXION

Developers

Free stack, honest disclosure, a real API over fake resources

14 core libraries and 8 developer tools in the fictional catalog — plus 13 live HTTP endpoints in this app that you can actually curl.

Reference

Demo API

No authentication. State is held in the server process, scoped to an anonymous cookie, and lost on restart. Quotas: 32 live instances and 8 endpoints per session.

GET/api/v1/catalogFull product catalog. Filter with ?family=, ?group=, ?q=
GET/api/v1/skusInstance types, hourly rates and region list
GET/api/v1/modelsServable models with token pricing and TTFT
GET/api/v1/statusSynthetic service and per-region capacity status
GET/api/v1/usageMetered spend for the current anonymous session
GET/api/v1/instancesList instances with live synthesized telemetry
POST/api/v1/instancesAllocate 1–16 instances of a SKU in a region
GET/api/v1/instances/{id}One instance, with hardware detail
DELETE/api/v1/instances/{id}Terminate and return the final cost
GET/api/v1/endpointsList serverless endpoints and their usage
POST/api/v1/endpointsDeploy a model endpoint (scale-to-zero)
DELETE/api/v1/endpoints/{id}Delete an endpoint
POST/api/v1/chat/completionsOpenAI-compatible shape; supports stream: true

curl

# no auth in the demo — an anonymous session cookie carries state
curl -sc jar http://localhost:3000/api/v1/skus | jq '.data[0]'

# allocate two 8-way M400 nodes with a job label
curl -sb jar -c jar http://localhost:3000/api/v1/instances \
  -H 'content-type: application/json' \
  -d '{"sku":"m400.8x","region":"us-west-2","count":2,"job":"axiom-70b-sft"}'

# watch telemetry and per-second cost climb
watch -n2 "curl -sb jar http://localhost:3000/api/v1/instances | jq -r \
  '.data[] | \"\(.id) \(.state) util=\(.metrics.utilization) \$\(.cost_usd)\"'"

# deploy an endpoint and call it
curl -sb jar http://localhost:3000/api/v1/endpoints \
  -H 'content-type: application/json' -d '{"model":"axiom-l-70b"}'

curl -sb jar http://localhost:3000/api/v1/chat/completions \
  -H 'content-type: application/json' \
  -d '{"model":"axiom-l-70b","messages":[{"role":"user","content":"hello"}]}' | jq -r \
  '.choices[0].message.content'

# tear everything down
curl -sb jar -X DELETE http://localhost:3000/api/v1/instances/i-xxxxxxxx

Python

from openai import OpenAI          # the shape is compatible; the model is not real

client = OpenAI(base_url="http://localhost:3000/api/v1", api_key="not-checked")

stream = client.chat.completions.create(
    model="axiom-l-70b",
    messages=[{"role": "user", "content": "what is this site?"}],
    stream=True,
)
for chunk in stream:
    print(chunk.choices[0].delta.content or "", end="")

What actually happens

The completions handler returns a fixed, self-identifying string and meters it. There is no model, no GPU and no inference. It exists so client code, streaming parsers and usage dashboards can be exercised against a stable target.

Catalog

Libraries and tools

View all →