Introducing Sprag: Unified Multimodal AI Inference

#Why we built Sprag

Running AI inference across multiple modalities today means juggling different providers, SDKs, billing systems, and latency profiles. You need one service for text generation, another for image understanding, a third for audio transcription, and yet another for video analysis. Each comes with its own authentication, rate limits, and failure modes. It's fragile, expensive, and slow to iterate on.

We built Sprag to collapse that complexity into a single managed API.

Under the hood, Sprag is powered by vLLM-Omni — our extension of the popular vLLM inference engine. What makes vLLM-Omni novel is that it serves text, image, audio, and video modalities through a single process with a unified KV-cache shared across all modalities. Traditional approaches stitch together separate model pipelines — one process for text, another for vision, another for audio — each with its own memory footprint and scheduling overhead. vLLM-Omni eliminates that overhead entirely by running everything in one engine with shared attention state.

The result: lower latency, higher throughput, and a dramatically simpler operational footprint. Sprag wraps this engine into a managed cloud API so you never have to think about GPU scheduling, model loading, or infrastructure scaling.

#What Sprag does

• Multimodal by default. Text generation, image understanding, audio transcription and synthesis, and video analysis — all through one consistent OpenAI-compatible API.
• Low latency at scale. Requests are routed to optimised vLLM-Omni instances with shared KV-cache, eliminating cold starts and redundant memory allocation across modalities.
• Simple integration. A single API key, a single endpoint pattern, and clear documentation. Go from zero to your first inference call in under five minutes.
• Usage-based pricing. Pay only for what you use. No reserved capacity, no minimums.

#Supported models at launch

We're launching with the following models, all served through the same API:

All models are available immediately. We'll be adding more as we validate performance and quality on our infrastructure.

#Making your first call

The Sprag API follows the OpenAI chat completions format. Here's a simple text generation request:

curl https://api.sprag.ai/v1/chat/completions \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer YOUR_API_KEY" \
  -d '{
    "model": "Qwen/Qwen3-Omni-30B-A3B-Instruct",
    "modalities": ["text"],
    "messages": [
      {"role": "user", "content": "Hello, how are you?"}
    ]
  }'

#Get started

Head over to our docs to grab your API key and make your first request. You can also try things out in the models page — no sign-up required for a quick test.

We're just getting started and would love your feedback. Reach out on our contact page or drop us a message — we read everything.