Open models, served fast,
and measured in public.
One endpoint for open-weight models. Change the base URL in the client you already use and keep everything else. Zero data retention: no conversation archive or training on your content; temporary inference caching is explained in our privacy policy. Every performance number on this page is dated and comes from a measurement we keep and will send you.
Why Creekside
Four reasons, each measured and dated, so you can check them before you switch.
250K tokens read in 52 seconds
Gemma 4 31B read a 250,000-token prompt from cold and began its answer in 52.2 s, through the public endpoint. Qwen3.8 27B read 200,000 tokens in 34.4 s (both 2026-11-08). The Qwen models take up to 1,048,576 tokens.
Under the OpenRouter median
Input and output, below the median of the providers OpenRouter lists for the same model: Qwen3.8 27B 20% and 24%, Qwen3.6 35B A3B 30% and 30%, Gemma 4 31B 20% and 10%, Muse Glimmer 30B 7% and 8% (read 2026-11-08).
Zero data retention, only
No conversation archive, content logging or training on your content, on every key. Temporary inference caching is explained in the privacy policy.
OpenAI-compatible
Change the base URL and keep your SDK. Anthropic's Messages API and OpenAI's Responses API are served too, so Claude Code and Codex CLI connect directly.
Speed you can check
Most providers publish a number and no method. These are what real callers got through the public endpoint, read from our own request records, with the reads kept alongside.
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.
curl
# Chat completion, non-streaming curl https://api.creeksidecompute.ai/v1/chat/completions \ -H "Authorization: Bearer $CREEKSIDE_API_KEY" \ -H "Content-Type: application/json" \ -d '{ "model": "qwen/qwen3.8-27b", "messages": [{"role": "user", "content": "Explain prefix caching."}], "max_tokens": 512 }'
Python
# pip install openai import os from openai import OpenAI client = OpenAI( base_url="https://api.creeksidecompute.ai/v1", api_key=os.environ["CREEKSIDE_API_KEY"], ) response = client.chat.completions.create( model="qwen/qwen3.8-27b", messages=[{"role": "user", "content": "Explain prefix caching."}], ) print(response.choices[0].message.content)
Node
// npm install openai import OpenAI from "openai"; const client = new OpenAI({ baseURL: "https://api.creeksidecompute.ai/v1", apiKey: process.env.CREEKSIDE_API_KEY, }); const response = await client.chat.completions.create({ model: "qwen/qwen3.8-27b", messages: [{ role: "user", content: "Explain prefix caching." }], }); console.log(response.choices[0].message.content);
Streaming
# Server-sent events. Usage counts arrive on the final chunk. stream = client.chat.completions.create( model="qwen/qwen3.8-27b", messages=[{"role": "user", "content": "Write a haiku."}], stream=True, stream_options={"include_usage": True}, ) for chunk in stream: if chunk.choices and chunk.choices[0].delta.content: print(chunk.choices[0].delta.content, end="") # Make the zero-retention guarantee explicit on any request: # x-creekside-retention: none
https://api.creeksidecompute.ai/v1
Drop-in replacement for the OpenAI base URL.
Authorization: Bearer
One key per project. To rotate, make a new key on your account page and revoke the old one once you have switched.
Chat, completions, models
Server-sent events for streaming. Usage counts on the final chunk.
Use it from your tools
Anything that speaks OpenAI's Chat Completions API connects with a base URL, a key and a model id. Claude Code and Codex CLI connect directly too, because we also serve Anthropic's Messages API and OpenAI's Responses API. Pick your tool for the lines to paste; its setup page has thinking, limits and fixes. No key yet? Chat with Qwen3.8 27B on Poe instead.
Claude Code
Creekside also serves Anthropic's Messages API, so Claude Code connects with no proxy. Set these, then run claude. The base URL has no /v1: Claude Code adds it.
export ANTHROPIC_BASE_URL=https://api.creeksidecompute.ai
export ANTHROPIC_AUTH_TOKEN=<your key>
export ANTHROPIC_MODEL=qwen/qwen3.8-27b
export ANTHROPIC_DEFAULT_HAIKU_MODEL=qwen/qwen3.8-27b
export ANTHROPIC_DEFAULT_SONNET_MODEL=qwen/qwen3.8-27b
export ANTHROPIC_DEFAULT_OPUS_MODEL=qwen/qwen3.8-27b
export CLAUDE_CODE_MAX_CONTEXT_TOKENS=1048576Codex CLI
Creekside also serves OpenAI's Responses API, so Codex connects with no proxy. Add this, with your key in CREEKSIDE_API_KEY.
model = "qwen/qwen3.8-27b" model_provider = "creekside" model_context_window = 1048576 web_search = "disabled" [model_providers.creekside] name = "Creekside Compute" base_url = "https://api.creeksidecompute.ai/v1" env_key = "CREEKSIDE_API_KEY" wire_api = "responses"
OpenCode
Add a provider, with your key in CREEKSIDE_API_KEY. The setup page adds the thinking variants.
{ "provider": { "creekside": {
"npm": "@ai-sdk/openai-compatible",
"options": {
"baseURL": "https://api.creeksidecompute.ai/v1",
"apiKey": "{env:CREEKSIDE_API_KEY}" },
"models": { "qwen/qwen3.8-27b": {
"limit": { "context": 1048576, "output": 32768 } } } } },
"model": "creekside/qwen/qwen3.8-27b" }pi
Add a provider to pi, the terminal coding agent, with your key in CREEKSIDE_API_KEY.
{ "providers": { "creekside": {
"baseUrl": "https://api.creeksidecompute.ai/v1",
"api": "openai-completions",
"apiKey": "$CREEKSIDE_API_KEY",
"compat": { "supportsDeveloperRole": false },
"models": [{ "id": "qwen/qwen3.8-27b", "reasoning": true,
"contextWindow": 1048576, "maxTokens": 65536 }] } } }Aider
Set the OpenAI-compatible variables, then name the model with Aider's openai/ prefix.
export OPENAI_API_BASE=https://api.creeksidecompute.ai/v1
export OPENAI_API_KEY=<your key>
aider --model openai/qwen/qwen3.8-27bCline
In Cline's settings, set the API Provider to OpenAI Compatible and fill in these fields.
Base URL https://api.creeksidecompute.ai/v1
API Key <your key>
Model ID qwen/qwen3.8-27b
Context Window Size 1048576
Image Support offKilo Code
Add a provider in Kilo's global config, with your key in CREEKSIDE_API_KEY.
{ "provider": { "creekside": {
"npm": "@ai-sdk/openai-compatible",
"options": {
"baseURL": "https://api.creeksidecompute.ai/v1",
"apiKey": "{env:CREEKSIDE_API_KEY}" },
"models": { "qwen/qwen3.8-27b": {
"limit": { "context": 1048576, "output": 32768 } } } } },
"model": "creekside/qwen/qwen3.8-27b" }Zed
Add an OpenAI-compatible provider for Zed's agent panel. Zed reads the key from CREEKSIDE_API_KEY.
{ "language_models": { "openai_compatible": { "Creekside": {
"api_url": "https://api.creeksidecompute.ai/v1",
"available_models": [{
"name": "qwen/qwen3.8-27b",
"max_tokens": 1048576,
"max_output_tokens": 32768 }] } } } }Hermes Agent
Add a named provider, with CREEKSIDE_API_KEY in ~/.hermes/.env.
providers:
creekside:
api: https://api.creeksidecompute.ai/v1
key_env: CREEKSIDE_API_KEY
models:
"qwen/qwen3.8-27b":
context_length: 1048576
model:
default: qwen/qwen3.8-27b
provider: custom:creeksideOpenClaw
Add a provider under models.providers, with CREEKSIDE_API_KEY in ~/.openclaw/.env, then openclaw models set creekside/qwen/qwen3.8-27b.
creekside: {
baseUrl: "https://api.creeksidecompute.ai/v1",
apiKey: "${CREEKSIDE_API_KEY}",
api: "openai-completions",
models: [{ id: "qwen/qwen3.8-27b", name: "Qwen3.8 27B",
reasoning: true, input: ["text"],
contextWindow: 1048576, maxTokens: 32768,
compat: { thinkingFormat: "qwen-chat-template" } }],
}OpenAI SDKs
The official SDKs work unchanged: give the client our base URL and your key.
import os
from openai import OpenAI
client = OpenAI(base_url="https://api.creeksidecompute.ai/v1",
api_key=os.environ["CREEKSIDE_API_KEY"])
reply = client.chat.completions.create(
model="qwen/qwen3.8-27b",
messages=[{"role": "user", "content": "Hello"}])No key needed
Chat with Qwen3.8 27B on Poe: open our Creekside profile. Poe bills you; nothing to set up.
NanoGPT and OpenRouter
Our routes are live on NanoGPT. OpenRouter onboarding is in progress; direct Creekside keys work now.
Models
Open weights only. We list a model when we can serve it well, and we publish the quantization we run it at, so you always know what you are getting.
| Model and id | Context | Quantization | Input | Output | Cached input | Speed | Below OpenRouter median |
|---|---|---|---|---|---|---|---|
| Qwen3.8 27B qwen/qwen3.8-27b | 1,048,576 | NVFP4 | $0.14 | $1.75 | $0.035 | 108 tok/s | 20% / 24% |
| Qwen3.6 35B A3B qwen/qwen3.6-35b-a3b | 1,048,576 | FP8 | $0.07 | $0.665 | $0.0175 | 280 tok/s | 30% / 30% |
| Gemma 4 31B google/gemma-4-31b-it | 262,144 | NVFP4 | $0.12 | $0.36 | $0.045 | 98 tok/s | 20% / 10% |
| Muse Glimmer 30B meta/muse-glimmer-30b | 131,072 | NVFP4 | $0.28 | $1.10 | $0.033 | 59 tok/s | 7% / 8% |
| Qwen 3.8 27B Uncensored qwen/qwen3.8-27b-uncensored | 1,048,576 | NVFP4 | $0.55 | $2.36 | $0.078 | 110 tok/s | Partners only |
coming soon Gemma 4 26B A4B, Qwen3 Coder Next, Qwen3.5 122B, GLM-5.3 and Kimi K3: open models from Google, Qwen, Z.ai and Moonshot AI.
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. Speed: output tokens a second for each caller at the median, on real traffic through the public endpoint on 2026-11-09, since 05:00 UTC for the NVFP4 Gemma and Qwen3.8 27B routes and over the 24 hours before 06:52 UTC for the others. Below OpenRouter median: how much lower our input / output prices are than the median price of the providers OpenRouter lists for the same model, read on 2026-11-08: 24 listings for Qwen3.8 27B, 11 for Qwen3.6 35B A3B, 16 for Gemma 4 31B and 1 for Muse Glimmer 30B. The uncensored route is available only to approved partners and has no comparable OpenRouter listing.
One tier, zero retention
Every price here is for zero data retention. Nothing is throttled and nothing is degraded.
Zero data retention
- No conversation archive or content logging
- Temporary inference caching, under the privacy policy
- Only usage metadata is kept, for billing
- The setting on every key you make
- Suitable where a data processing agreement is required
What we do with your traffic
Short version: as little as possible, and it is enforced by how the service is built, not promised in prose.
Zero data retention
No conversation archive or content logging. Temporary inference caches avoid repeated prompt processing; the privacy policy explains their use. Accounting records carry token counts and a request id, never text.
The header makes it certain
Send x-creekside-retention: none and that request is never stored, whatever the key says. The header always wins.
Expires in 13 months
Usage records carry no content, and they expire automatically. They serve billing, abuse prevention and capacity planning.
Only to serve your requests
Your content and temporary inference state are not used for training, sale or licensing.
Less repeated work
Reusable model state can make follow-ups faster and cheaper. Cache storage remains subject to your customer or marketplace terms; customer-content disk caching is not currently enabled.
Ask for the raw runs
Every speed number here comes from a benchmark run or a read of our request records that we keep, including the ones that went badly. Ask and we send it.
Questions, answered first
Is the API OpenAI-compatible?
Yes. Point any OpenAI SDK or OpenAI-compatible client at https://api.creeksidecompute.ai/v1 with your key. Chat and completions, streaming, tool calling, structured output and usage accounting all work as your client expects.
Which models do you serve, and what do they cost?
Four models are self-serve: Qwen3.8 27B at $0.14 input, $1.75 output and $0.035 cached input per million tokens; Qwen3.6 35B A3B at $0.07, $0.665 and $0.0175; Gemma 4 31B at $0.12, $0.36 and $0.045; and Muse Glimmer 30B at $0.28, $1.10 and $0.033. Every price is for zero data retention. Qwen3.8 27B Uncensored is also served at $0.55, $2.36 and $0.078 for approved partners. The catalogue at GET /v1/models is authoritative.
How fast is it?
On real traffic through the public endpoint on 2026-11-09, each caller got, at the median, 98 tokens a second on Gemma 4 31B, 108 on Qwen3.8 27B, 280 on Qwen3.6 35B A3B and 59 on Muse Glimmer 30B, with the first token in 0.43 to 0.69 s on Gemma and the Qwen models. Long prompts read fast: on 2026-11-08, Gemma 4 31B began answering a 250,000-token prompt from cold in 52 s, and Qwen3.8 27B a 200,000-token prompt in 34 s. Over the 7 days to 2026-11-09, 99.94% of requests completed in full. Every number on this page carries the date it was measured.
Do you store my prompts or train on them?
Zero data retention is the only tier we offer: no conversation archive, content logging, training, sale or licensing. Temporary inference caching is disclosed in our privacy policy and remains subject to customer and marketplace terms. Send x-creekside-retention: none on any request to make sure it is never stored, whatever the key says.
What context window do you support?
1,048,576 tokens of context on the Qwen models, 262,144 on Gemma 4 31B and 131,072 on Muse Glimmer 30B, with up to 65,536 output tokens per request on every model.
How do I get a key?
Make an account at creeksidecompute.ai/account: sign in with your email and make your keys there. Each request is paid from prepaid credit: top up by card from $10, or redeem a promotional code. Larger users can ask for a direct agreement, invoiced monthly.
Can I use it from coding agents and chat apps?
Yes. Claude Code and Codex CLI connect directly, because we also serve Anthropic's Messages API and OpenAI's Responses API. Anything that speaks OpenAI's Chat Completions API with a custom base URL works too: OpenCode, pi, Aider, Cline, Kilo Code, Zed, Hermes Agent, OpenClaw and the OpenAI SDKs. Each has a setup page, linked from Connect above.
What is Creekside Compute?
An independent inference provider for open-weight models, launched in 2026. It sells tokens through its own API and through marketplaces, publishes the measurements behind every speed claim, and offers zero data retention only, with temporary inference caching disclosed in its privacy policy.
Start
Make an account and a key
Sign in with your email and make a key. Each request is paid from prepaid credit: top up by card from $10, or redeem a promotional code. Larger users can ask for a direct agreement, invoiced monthly.
Change the base URL
One line in your client. Keep your prompts, your SDK and your tooling.
Compare and decide
Run your own load. If we are not faster or cheaper for your shape of traffic, we would rather you knew. Testing for a product or a platform? Write to us about test credit.