01
Byte-for-byte streaming
Tokens leave the moment they are generated, never accumulated and re-serialized. Time to first token stays what the GPU produced.
Open models under every listed price
Nyx serves open-weight models on tuned GPU capacity behind one OpenAI-compatible endpoint. Byte-for-byte streaming, 71µs of routing overhead at p50, and every model priced under its cheapest listing.
Powering inference for
A model library that stays priced to win
Frontier open-weight models, serverless by default. Prices in USD per million tokens, published live at /v1/models, and always under the cheapest listing for the model.
2.1B+tokens served monthly
99.95%uptime over the last 30 days
71µsrouting overhead at p50
<24hto a dedicated endpoint
Measured against the market
Weekly benchmark runs on gpt-oss-120b, same prompts, same
day, against the three largest serverless providers for the model.
Throughput
output tokens per second, higher is better
Output price
USD per million tokens, lower is better
Streaming enabled, 1K-token prompts, 512-token completions, us-east clients.
Engineered for the serving path
Every request rides a serving layer measured in microseconds, in front of tuned GPU capacity. The properties below are measured on every build, and a release that regresses one does not ship.
02
Failover before the first token
Capacity fails over transparently before generation starts. A stream is never spliced together from two completions.
03
Backpressure, not brownouts
Past declared capacity Nyx answers immediately instead of queueing requests into timeouts, so your retries land somewhere useful.
04
Overhead in microseconds
The serving layer adds 71µs at p50 and 235µs at p99, published and regression-tested. Latency budgets stay with your model.
See every token you serve
Usage, spend, latency, and per-request logs in the console, live as the tokens stream. Export everything; it is your traffic.
Dedicated capacity for production workloads
Reserved GPU capacity behind the same endpoint, tuned to your model and traffic and isolated from public load. Set up in under 24 hours, backed by an SLA.
Explore dedicatedShip in an afternoon
Point any OpenAI SDK at https://nyxprovider.com/v1 and go.
Streaming is byte-for-byte; responses match the schema you already
parse.
terminal
curl https://nyxprovider.com/v1/chat/completions \
-H "Content-Type: application/json" \
-d '{
"model": "gpt-oss-120b",
"stream": true,
"messages": [{"role": "user", "content": "Hello"}]
}'