# Running 5 AI agents on one 5090 — how would you handle the VRAM? — retained record

> Retained reddit record from reddit:homelab.

Canonical: https://highsignal.app/data/records/e90f373ebcf947f1

[← sources](https://highsignal.app/data)


source: [reddit](https://highsignal.app/data/reddit) (reddit:homelab)  published 2026-10-19 15:44 ingested 2026-10-20 02:39

[original source ↗](https://www.reddit.com/r/homelab/comments/1xa0dj3/running_5_ai_agents_on_one_5090_how_would_you/)

## retained text



```
Been planning a local multi-agent setup on my gaming PC and the VRAM math is the part I keep going back and forth on.

RTX 5090, 32GB. Windows. Plan is 5 agents: one coordinator (Qwen3:32b) plus 4 workers (14b/8b mix). Not all loaded at once, coordinator dispatches and workers spin up as needed. Everything through Ollama.

The math says a 32b at q4\_k\_m is around 18-20GB with context, leaving room for one worker. But Windows eats some VRAM for display, and context length can blow up fast when agents start pulling in memories.

Anyone running multiple LLMs on a single card? How do you manage it? Swap models in and out, or just keep the big one loaded and accept the wait? Any Ollama settings that actually help?
```



document fetched 2026-10-20 00:36 · canonical https://reddit.com/r/homelab/comments/1xa0dj3/running_5_ai_agents_on_one_5090_how_would_you
