Replicate + Meridian for hybrid open/closed models

Run open-source models on Replicate's GPU cloud while routing sensitive or proprietary workloads through Meridian's private inference. One API surface, zero model lock-in.

Your App

→

Meridian Router

Replicate (Llama, SDXL)

Meridian Private (your fine-tunes)

Model routing happens at the edge — single endpoint, dual backends

Quickstart

// Install both SDKs
npm install meridian replicate

// Configure
import { Meridian } from 'meridian';
import Replicate from 'replicate';

const meridian = new Meridian({ apiKey: process.env.MERIDIAN_KEY });
const replicate = new Replicate({ auth: process.env.REPLICATE_TOKEN });

// Route based on model prefix
async function infer(model: string, input: Record<string, unknown>) {
  if (model.startsWith('replicate/')) {
    return replicate.run(model.replace('replicate/', ''), { input });
  }
  return meridian.completions.create({ model, ...input });
}

Routing Table

Prefix	Backend	Use Case
replicate/	Replicate	Open models, image gen, public benchmarks
meridian/	Meridian Private	Fine-tuned models, proprietary data, compliance workloads
(no prefix)	Meridian Default	Fallback — routes to cheapest available backend

Why Hybrid?

Cost Efficiency

Route bulk inference to Replicate's commodity GPU pricing. Reserve Meridian private nodes for latency-sensitive or compliance-bound requests. Typical savings: 40-60% vs all-private.

Data Sovereignty

Keep fine-tuned weights and proprietary prompts inside Meridian's isolated tenants. Replicate never sees your model internals or sensitive input schemas.

Next Steps

Meridian Quickstart→Deploying Private Models→Replicate Docs ↗→