Every open model.
One key.

Qwen, DeepSeek, Kimi, GLM and MiniMax behind one OpenAI-compatible key. Point your client at us; everything else keeps working.

Get an API key Read the docs

RelayingQwenDeepSeekMoonshotZ.AIMiniMaxWanHunyuan53 models · one key

The relay

One key in front.
Every model behind.

deeprelay sits between your client and the open-model frontier. Requests stream in on one OpenAI-compatible key, the relay routes each one to the model you named, and tokens stream back. The stream above is the product drawn literally.

The relay · one wire

your app ──▶ api.deeprelay.ai ──▶ deepseek-v4
                            ├──▶ glm-5.3
                            ├──▶ kimi-k3
                            ├──▶ qwen-image
                            ├──▶ wan-2.2
                            └──▶ …38 models, one key

Chat

The chat surface, byte for byte.

Streaming, tool calls, JSON mode and the usage block behave the way your client already expects, including reasoning tokens, surfaced in usage instead of silently swallowed.

SSE frames your parser already handles
Tool calls and structured output
Reasoning tokens reported in usage

Browse chat models →

Python · streaming

stream = client.chat.completions.create(
    model="deepseek-v4",
    messages=[{"role": "user", "content": q}],
    tools=tools,
    stream=True,
)
for chunk in stream:
    delta = chunk.choices[0].delta

Image

Images on the same key.

The OpenAI images shape, pointed at the open image models: one request, a finished image, accounted inside the plan like everything else.

Standard /v1/images/generations
Per-image accounting, inside the plan
The response your client already parses

Browse image models →

curl · images

curl https://api.deeprelay.ai/v1/images/generations \
  -H "Authorization: Bearer $DEEPRELAY_API_KEY" \
  -d '{
    "model": "qwen-image",
    "prompt": "a watercolor fox",
    "size": "1024x1024"
  }'

Video

Video is a job, not a timeout.

Generation runs as an async job: submit, poll, download. Long renders never hold a connection open, and seconds of video are accounted the same way as tokens, inside the plan.

POST /v1/videos → job id
Poll status, then fetch /content
Per-second accounting, same key

Browse video models →

curl · video jobs

curl https://api.deeprelay.ai/v1/videos \
  -d '{"model": "wan-2.2", "prompt": "..."}'
# → {"id": "vid_…", "status": "queued"}

curl https://api.deeprelay.ai/v1/videos/vid_…/content
# → the finished render

The catalog