Most WordPress AI plugins ship with an OpenAI key field and a model dropdown. Point them at a self-hosted vLLM endpoint running on the RTX 5060 Ti 16GB at our UK dedicated GPU hosting and you replace per-call token charges with a flat monthly line item, while keeping drafts, prompts and customer comments on UK infrastructure. The card gives 4608 Blackwell CUDA cores, 16 GB of GDDR7, 448 GB/s of bandwidth and native FP8, which is enough to back a busy multi-author WordPress site without queueing.
Contents
- Compatible plugins and where to paste the URL
- Endpoint configuration
- Latency and throughput numbers
- Cost comparison vs OpenAI
- Features you can drive on a single card
- Privacy and GDPR
Compatible plugins
Any plugin that exposes a “custom OpenAI base URL” or “proxy endpoint” field works. Tested configurations include AI Engine (Jordy Meow), AIKit, BetterAI, GetGenie, Bertha AI, Rank Math Content AI, CodeWP and a handful of WooCommerce description generators. All of them speak the OpenAI Chat Completions schema, which vLLM serves by default, so there is no code to modify.
Configuration
Boot vLLM on the 5060 Ti, open your plugin settings, and paste three values.
| Plugin field | Value |
|---|---|
| API base URL | https://llm.yoursite.co.uk/v1 |
| API key | Any non-empty bearer token (validated by your reverse proxy) |
| Model name | meta-llama/Meta-Llama-3.1-8B-Instruct-FP8 |
| Max tokens | 1024-2048 for blog drafts, 128 for titles |
| Temperature | 0.3 for SEO, 0.7 for creative |
Put Nginx or Cloudflare in front with a static bearer token and a per-IP rate limit. See our FP8 Llama deployment guide for the vLLM command line.
Latency and throughput
| Task | Model | Tokens | Time on 5060 Ti |
|---|---|---|---|
| Meta title + description | Phi-3 mini FP8 | 80 out | 0.3 s |
| 600-word blog draft | Llama 3.1 8B FP8 | 900 out | 8.0 s |
| Comment moderation verdict | Phi-3 mini FP8 | 20 out | 0.1 s |
| Product description (WooCommerce) | Mistral 7B FP8 | 250 out | 2.1 s |
| Bulk tag suggestion (batch 20) | Llama 3.1 8B FP8 | 2000 out total | 2.8 s |
Single-stream Llama 3.1 8B FP8 runs at about 112 t/s and aggregate throughput across concurrent plugin calls reaches roughly 720 t/s. Around 16 editors can trigger long-form generation simultaneously without visible queueing.
Cost comparison
| Monthly usage | OpenAI GPT-4o-mini | Self-hosted 5060 Ti |
|---|---|---|
| 500 blog drafts + 10k moderations | ~£180 | Flat £300 (includes headroom) |
| 5k WooCommerce descriptions | ~£90 | Same box, same price |
| 50k SEO meta generations | ~£150 | Same box, same price |
The break-even is around the first heavy-use month; after that every additional task is effectively free.
Features
- Long-form post drafts from outline or URL
- SEO titles, meta descriptions and schema JSON
- Auto-tagging and category suggestions
- Comment and review moderation with reasoned verdicts
- Translation of posts into additional languages via Qwen 2.5 14B AWQ
- Featured image generation using SDXL Lightning on the same card
Privacy
Draft content, unpublished commercial posts and reader comments never leave your rack. There is no third-party sub-processor to declare in your privacy policy and no US data transfer clause to negotiate. For sites under ICO scrutiny this is often the deciding factor over performance.
Swap your OpenAI key for a private endpoint
Self-hosted vLLM for WordPress AI plugins. UK dedicated hosting.
Order the RTX 5060 Ti 16GBSee also: AI-powered CMS, e-commerce AI, Llama 3 8B benchmark, FP8 Llama deployment, internal tooling.
