Zero-retention inference · sign-up open
Your data is not our data.
An OpenAI-compatible endpoint. Nothing you send is written to disk.
or read the quickstartPOST /v1/chat/completionsbase_urlhttps://api.inferway.ai/v1stream true
01Privacy
Private by construction.
There is no logging stage to opt out of — persistence doesn't exist in the path.
01
Request sent
TLS-encrypted, straight to a dedicated endpoint.
02
Routed
No shared queue in the path.
03
Inference in GPU memory
Paged KV, never written to disk.
04
Destroyed
Erased on completion; only metadata remains.
02Model & pricing
Model and pricing.
from catalogThe current model
Live262K ctxNVFP4max out 32,768Zero retention
Input vs. cache hit
Input$0.40
Cache hit$0.06
A cache hit costs about 15% of a normal input token.Input
$0.40/ 1M
Output
$2.80/ 1M
Cache hit
$0.06/ 1M
Per token, metered from metadata. No seats, no minimums.
Rough monthly cost$744/ month
20,000 calls / day1,000 in · 300 out
03Benchmarks
Numbers are being re-measured.
Measuring nowFirst-token latencypendingMeasured at fixed concurrency
Output throughputpendingTokens per second, single stream
90-day uptimependingPer service, from the status page
Throughput and latency are being re-measured on the current hardware and model. Until that finishes, this section stays empty — we publish measurements with hardware, concurrency and timestamp attached, or we publish nothing.
04Integrate
Keep your SDK.
Create an accountConsole
no credit cardPoint base_url at Inferway
https://api.inferway.ai/v1Run the streaming smoke test
stream=trueFree use: 60/min · 120/hour · 1000/day
Long outputs require stream=true
Streaming
curl https://api.inferway.ai/v1/chat/completions \
-H "Authorization: Bearer $INFERWAY_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "inferway/qwen3-8-27b",
"stream": true,
"messages": [{"role": "user", "content": "Say hello in one short sentence."}]
}'Copy setup for your AI assistant
Paste into Claude Code or Cursor. Key stays in your .env, never in the prompt.05Playground
Try it without an account.
No key needed · free use: 60/min · 120/hour · 1000/dayIW
Ask about data retention, pricing, streaming, or limits — or just chat.
Questions
No. Content is never written to disk — inference runs entirely in GPU memory and is destroyed on completion. We retain request metadata only (token counts, timestamps, latency, routing, HTTP status) for up to 90 days, for billing and reliability.
Input tokens that hit the cache are billed at $0.06 per 1M, against $0.40 for normal input and $2.80 for output. A cache hit costs about 15% of a normal input token, so long shared system prompts and retrieval contexts benefit most. All three prices come from the model catalog.
Non-streaming requests pass through a CDN with roughly a 100-second first-byte timeout. A long output returns no first byte until generation finishes, so it hits that timeout. With stream=true, bytes start arriving with the first token. Non-streaming is fine for short completions.
You pay per token, metered from request metadata, never from your request content. Pricing and context limits are published on the model catalog and generated from a single catalog source.
Longer answers: Docs · Privacy · Transparency