OpenAI-compatible inference
Drop-in replacement for the OpenAI API. Your requests run on distributed GPUs with signed receipts proving exactly what model ran, on which node, with what result.
Python — two lines to switch from OpenAI
from openai import OpenAI
client = OpenAI(
base_url="https://ryvion-hub.fly.dev/v1",
api_key="YOUR_RYVION_API_KEY",
)
response = client.chat.completions.create(
model="phi-4",
messages=[{"role": "user", "content": "Hello!"}],
stream=True,
)
for chunk in response:
print(chunk.choices[0].delta.content, end="")Built for real infrastructure work
The problem
Every inference API is a black box. You send a prompt, you get a response. You trust that the model you asked for is the model that ran. Ryvion changes that.
01
Other APIs say they ran GPT-4 or Llama 70B. You have no way to prove it. For compliance-sensitive workloads, this is unacceptable.
02
One provider, one data center, one point of failure. If they go down, change pricing, or change terms — you have no alternative.
03
Your prompts go wherever the provider decides. No control over which country processes your data.
How it works
Same API you already use. New guarantees you can't get anywhere else.
Step 1
Point the OpenAI SDK at Ryvion. One line change. Your existing code, prompts, and tools keep working.
Step 2
Your requests route to real GPU nodes across the network. Streaming and non-streaming, chat and embeddings.
Step 3
Every inference run produces a cryptographic receipt — proof of which model ran, on which node, with what result. Audit-ready by default.
Available models
Seven models across four modalities. Only models that online nodes can actually serve.
| Model | Type | Parameters | Speed |
|---|---|---|---|
| phi-4 | Chat | 14B | Fast |
| ryvion-llama-3.2-3b | Chat | 3B | Very fast |
| tinyllama | Chat | 1.1B | Instant |
| sdxl-turbo | Image | — | ~4s / image |
| stable-diffusion-xl | Image | — | ~15s / image |
| whisper-1 | Audio | base | Real-time |
| whisper-large | Audio | large-v3 | Near real-time |
Chat models via llama.cpp, images via Stable Diffusion, audio via Whisper — all on real GPU nodes. More models added as operators bring hardware online.
Verified inference
Not a log. A cryptographic proof tied to the node that ran your inference.
curl — try it now
curl https://ryvion-hub.fly.dev/v1/chat/completions \
-H "Authorization: Bearer YOUR_RYVION_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "phi-4",
"messages": [{"role": "user", "content": "Hello!"}],
"stream": true
}'What the receipt proves
This is what makes Ryvion different from every other inference API. Verifiable compute, not blind trust.
Why Ryvion
Other APIs resell GPU time. Ryvion is a verified compute network where every response is provably real.
Dimension
Ryvion
OpenAI-compatible. Change base_url in one line, keep your existing code.
Traditional API
Custom SDKs, proprietary formats, migration effort for each provider.
Dimension
Ryvion
Signed receipts per inference — cryptographic proof of model, node, and result.
Traditional API
Trust the provider. No proof of what model actually ran your request.
Dimension
Ryvion
Distributed GPU network. No single point of failure. Operators compete on price.
Traditional API
Centralized data centers. Provider sets the price. You hope they stay online.
Dimension
Ryvion
Jurisdiction-aware routing. Your data stays where you specify.
Traditional API
Your prompts go wherever the provider decides. No control.
Dimension
Ryvion
Per-token, transparent. Network competition drives prices down over time.
Traditional API
Per-token but opaque margins. Prices set by one company.
Use the OpenAI SDK with one line changed. Get verified inference with signed receipts, per-token pricing, and scoped API keys.
Contribute GPU capacity to the network. Open-source node agent (Apache 2.0). Earn from every verified inference your node completes.
FAQ
Common questions about Ryvion.
Ryvion is OpenAI-compatible — same API format, same SDKs. The difference: your requests run on a distributed GPU network with signed receipts proving exactly what model ran on which node. No other API gives you that level of verifiability.
Change your base URL to https://ryvion-hub.fly.dev/v1 and use your Ryvion API key. That's it. Your existing code, prompts, and tools keep working.
Chat: Phi-4 (14B), Llama 3.2 3B, TinyLlama (1.1B). Images: SDXL Turbo, Stable Diffusion XL. Audio: Whisper base and large-v3. Embeddings: all-MiniLM-L6-v2. More models are added as operators bring hardware online.
Every inference request produces a cryptographic receipt — proof that a specific model ran on a specific node and produced a specific result. This matters for compliance, audit trails, and trust.
Chat and embeddings: $0.06 per million tokens. Image generation: $0.03 per image. Audio transcription: $0.006 per minute. Every new account gets $10 in free credits — no credit card required. Additional credits via Stripe. Failed jobs are automatically refunded.