OpenAI-compatible · streaming · tool calling
Open models, served fast,
and measured in public.
One endpoint for open-weight models, served on dedicated accelerated compute. Change the base URL in the client you already use and keep everything else. Nothing is stored by default, so on a zero-retention key there is no copy of your traffic to train on or hand over. Every performance claim on this page traces to a benchmark run you can read.
Measured end to end through the public endpoint, Qwen3.8 27B
2026-09-22
Time to first token
517 ms
4K prompt, single stream
Per-stream decode
204 tok/s
what one caller sees
Node throughput
4,173 tok/s
sustained under load
Context
262K
on long-context lanes
01
Speed you can check
Most providers publish a number and no method. These are measured end to end through the same public endpoint you would call, on a single node, with the raw results kept alongside.
| Concurrent requests | Time to first token | Per-stream decode | Total throughput |
|---|
| 1 | 517 ms | 204 tok/s | 1,437 tok/s |
| 4 | 1.0 s | 149 tok/s | 3,707 tok/s |
| 16 sustained | 12.6 s | 143 tok/s | 4,173 tok/s |
| 8 | 4.5 s | 143 tok/s | 4,166 tok/s |
We size each node to the concurrency it serves well and add capacity rather than overloading it. That is why your time to first token stays flat as our traffic grows, instead of climbing whenever someone else gets busy.
02
Connect in one line
The API is OpenAI-compatible. Point your existing client at our base URL. Streaming, tool calling and usage accounting all behave the way your SDK already expects.
Base URL
https://api.creeksidecompute.ai/v1
Drop-in replacement for the OpenAI base URL.
Auth
Authorization: Bearer
One key per project. Rotate from the dashboard without downtime.
Compatibility
Chat, completions, models
Server-sent events for streaming. Usage counts on the final chunk.
03
Models
Open weights only. We list a model when we can serve it well, and we publish the precision we run it at, so you always know what you are getting.
| Model | Context | Precision | Input | Output | Cached input |
|---|
| Qwen3.8 27B | 262,144 | FP4 | $0.15 | $1.80 | $0.035 |
| Gemma 4 26B-A4B soon | 262,144 | FP8 | — | — | — |
| gpt-oss 20B soon | 131,072 | MXFP4 | — | — | — |
Prices are US dollars per million tokens. Cached input applies automatically when your prompt shares a prefix with a recent one, which is most of the time for agents and chat.
04
Two tiers, one difference
The only thing that separates them is whether we keep your request content. Nothing is throttled and nothing is degraded. The discount on the retained tier is what we pay you for the data rights.
Zero retention
$1.80 / M output
- Request and response content is never written to disk
- Only usage metadata is kept, for billing
- The default for every new key
- Suitable where a data processing agreement is required
Retained · 25% off
$1.35 / M output
- Content stored, then deleted automatically within 90 days
- You grant us rights to the stored content, including licensing it
- Opt in per key or per request with a header
- Useful when you want us to debug a bad response with you
Retention is off unless you turn it on. Send x-creekside-retention: none on any single request and that request is never stored, whatever the key says. The header always wins.
05
What we do with your traffic
Short version: as little as possible, and we enforce it in infrastructure rather than promising it in prose.
Default
Nothing is stored
Content is held in memory for the length of the request and then dropped. The accounting record carries token counts and a request id, never text.
If you opt in
Deleted in 90 days
Enforced by an object lifecycle rule on the bucket itself, so it happens whether or not our code remembers to.
Metadata
Expires in 13 months
Usage records expire by partition policy on the warehouse. Ask us to delete a key's history and we run one tool.
Zero-retention keys
Nothing exists to use
Content is never written to disk, so there is no copy to train on, license, hand over or lose. The guarantee is structural, not a promise about our intentions.
Where it runs
Isolated, ephemeral nodes
Serving nodes hold no long-term storage, accept no inbound connections, and are wiped when released.
Transparency
Read the measurements
Every speed claim here links to the benchmark run that produced it, including the ones that went badly.
06
Start
Step one
Get a key
Self-serve, no call. Free credit to benchmark us against whatever you use now.
Step two
Change the base URL
One line in your client. Keep your prompts, your SDK and your tooling.
Step three
Compare and decide
Run your own load. If we are not faster or cheaper for your shape of traffic, we would rather you knew.