Cynic Inference: an OpenAI-compatible API for open-weight models

Rolling out

We are bringing the first three models online now. The prices below are the prices we bill. For an API key, or to talk about volume, email [email protected].

Ask for a key See the benchmarks

Step-3.7-Flash Rolling out

By StepFun, under the Apache-2.0 license, served from official NVFP4 weights.

$0.19: Input
$0.038: Cached input
$1.09: Output

Model ID: stepfun/step-3.7-flash
Context window: 262,144 tokens
Precision: FP4

MiMo-V2.5 Rolling out

By Xiaomi, under the MIT license, served from official FP8 weights.

$0.30: Input
$0.06: Cached input
$1.50: Output

Model ID: xiaomi/mimo-v2.5
Context window: 262,144 tokens
Precision: FP8

Qwen3.6-35B-A3B Rolling out

By Qwen, under the Apache-2.0 license, served from official FP8 weights.

$0.10: Input
$0.025: Cached input
$0.90: Output

Model ID: qwen/qwen3.6-35b-a3b
Context window: 262,144 tokens
Precision: FP8

Need another model?

We can host almost any open-weight text model, from small coding models to the newest large releases, as long as its license allows serving it as an API. Over a hundred are on our quick-request list, most running within about 30 minutes; for anything else, tell us the model, the volume you expect and the context length you need.

Quick request list Request another model

Zero data retention: Prompts and completions are never logged or stored. We keep only request metadata (time, model, token counts, latency, status and which API key) for billing, capacity and abuse handling. Data policy
No training on your data: We do not train or fine-tune any model on what you send or receive, and we do not sell or share it.
Prices per token: Input, cached input and output, each with its own price on this page. Nothing else is billed.
Cached input costs less: Input the model has just seen, such as a shared system prompt or the earlier turns of a conversation, is served from GPU memory, billed at the cached price and reported in usage.prompt_tokens_details.cached_tokens.
Full context windows: Each model runs with the whole context window it was released with: 262,144 tokens for all of them.
Busy, not queued: We size each server so every stream keeps its speed. When a model is at capacity we answer 429 at once instead of holding your request in a queue, so your client or router can retry elsewhere without waiting.

Point your client at the base URL, send your API key as a bearer token and name a model by its ID. Responses stream as server-sent events, and the usage block counts input, cached input and output tokens.

Base URL: https://api.cynic-hosting.com/v1
Chat: POST /chat/completions, streaming and tool calls
Responses: POST /responses, stateless, for Codex and the OpenAI Agents SDK
Anthropic: POST /messages, for Claude Code and Anthropic clients (key in x-api-key)
Models: GET /models, with each model's prices and context window
Busy: 429 with a Retry-After header

curl https://api.cynic-hosting.com/v1/chat/completions \
  -H "Authorization: Bearer $CYNIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "qwen/qwen3.6-35b-a3b",
    "messages": [{"role": "user", "content": "Write a haiku about rust."}],
    "stream": true
  }'

The GPUs

The models run on dedicated NVIDIA data-center GPUs that we rent from a cloud GPU provider, in its secure cloud. The provider is named in our data policy. We measure each model on the hardware before we offer it, and publish what we measured.

Benchmarks

The gateway

Your requests reach the GPUs through our own API gateway, which runs on a server we own. It checks your key, counts tokens for billing and routes each request to a GPU with room for it. It does not write prompts or completions anywhere. Cynic Hosting LLC is a company in the United States.

Privacy policy

Policies

Terms of service Privacy policy Data policy

Cynic Inference

Rolling out

Models