Hardware requirements for a local LLM server with MeigaHub

Profile	Expected usage	GPU	VRAM	RAM	Recommended model	Power cost*
Just me Freelance · 1 user	~50k tokens/day Occasional chat, 1-2 automations	RTX 3060 / 4060	12 GB	16 GB DDR4	Qwen 2.5 7B Llama 3.1 8B	~€8/mo
Small business Most recommended	~500k tokens/day 5 users, campaigns, active leads	RTX 4070 Super / 4070 Ti	12-16 GB	32 GB DDR4/DDR5	Qwen 2.5 14B Mistral Nemo 12B	~€15/mo
Mid-sized business 20 users	~2M tokens/day Sales team + heavy automation	RTX 4090	24 GB	64 GB DDR5	Qwen 2.5 32B Llama 3.3 instruct	~€25/mo
Agency / large team 50+ users	~10M tokens/day 24/7 production, blog autopilot, multi-tenant	RTX A6000 · 2×RTX 4090	48 GB	128 GB DDR5	Llama 3.3 70B Qwen 2.5 72B	~€60/mo

* Estimated at €0.15/kWh and average daily usage. It varies with your electricity tariff, cooling and real load.

FAQ

Can I start without a dedicated GPU?

Yes. MeigaHub lets you use cloud APIs (OpenAI, Mistral, Groq) with no server of your own. You can also run small models (3-4B) on plain CPU for testing. A local LLM server only becomes mandatory when you want full privacy and zero cost per token.

Operating system: Windows or Linux?

MeigaHub-server runs on both. Windows 10/11 with CUDA is the quickest to get going. Linux Ubuntu 22.04+ is the most stable for 24/7 production. Either way, the connection to MeigaHub goes over Tailscale.

Do AMD or Intel work too?

We recommend NVIDIA (CUDA) for compatibility and performance. AMD ROCm works on Linux with llama.cpp models, but support is more fragile. Intel Arc is not yet production-ready for large LLMs. If you are going to invest, go NVIDIA.

Can I use several GPUs in parallel?

Yes. MeigaHub-server supports tensor parallel and model split across 2-4 GPUs. That lets you load Llama 70B on 2×RTX 4090 (24 GB each). You can also register several servers in MeigaHub and the routing balances them automatically by priority and load.

Does the server need internet? An open port?

You only need outbound internet (egress) — there is no router port to open. MeigaHub-server joins MeigaHub's private tailnet over Tailscale, so you do not even need a fixed IP. It works behind NAT and CGNAT.

What if my PC shuts down? Do I lose the service?

If you set cloud APIs as a fallback in your MeigaHub account, the agent keeps working without your local server (charging per token). If you only rely on the local one, tasks are retried automatically once it is back online. For serious production we recommend a dedicated mini-PC or a desktop server that stays on.

What hardware do I need
for my MeigaHub server?

FAQ

Want a preconfigured server?

What hardware do I needfor my MeigaHub server?

FAQ

Want a preconfigured server?

What hardware do I need
for my MeigaHub server?