Share one model across two Macs

Pro

You’ll get: one model name, @pool/qwen-27b, that your apps call. Behind it, the same local model runs on 2 Macs, MS1 and MM1. Nexus sends each request to the Mac that can answer it soonest. When you restart one Mac, requests go to the other.

You need:

• 2 Macs running Nexus, both in your fleet. Set up the pool on MS1.
• A local model on MS1, such as a Qwen coding model.
• A Nexus API key from Connections for your app.

Steps

1. On MS1, open Smart Balancer → Nodes and click Add node.
2. Next to MM1, click Add.
3. Put the same model on MM1. Either copy it from MS1, or open Deployments, click New deployment, pick MM1 and the model, and click Deploy model.
4. Open Pools and click New pool.
5. Enter the Name “Qwen 27B” and the Pool slug qwen-27b.
6. Set Compatibility mode to Exact replicas.
7. Click Add target. Pick Machine MS1 and your model.
8. Click Add target again. Pick Machine MM1 and the same model.
9. Click Save changes.

If the Macs run different files of the model, such as MLX on MS1 and a GGUF file on MM1, pick Equivalent generation instead.

If Nexus asks Same model on every machine?, confirm only if both Macs have the same files, as after a copy or a deployment.

Now point your app at the pool, on MS1’s address:

curl http://127.0.0.1:3939/v1/chat/completions \
  -H "Authorization: Bearer msty_nexus_key_here" \
  -H "Content-Type: application/json" \
  -d '{"model": "@pool/qwen-27b", "messages": [{"role": "user", "content": "Say hello."}]}'

Change the key to yours. An app on another computer uses MS1’s network address instead of 127.0.0.1.

Check it worked

1. On Pools, open the pool’s menu (…) and click Failure drill.
2. Nexus sends test requests, pauses one Mac, then brings it back. Each step, Before, One machine down and After recovery, shows how many requests were served.
3. Open Request history. Each request names the Mac that answered it.

To do it by hand: on Nodes, open MM1’s menu (…) and click Pause. Live status shows Paused — safe to restart. Send the curl again: it still gets a reply, and Request history shows MS1 answered. Click Resume when you’re done.

The drill uses real requests

Failure drill sends real requests to the models in the pool. With local models that costs nothing, but it keeps both Macs busy for a short time.

Pause a Mac on Nodes before you restart it or change its model. More in Smart Balancer.