Goldnat Plugins
← Back to documentation home

Choosing the AI for the XenForo Assistant — API or Your Own Server

Three ways to run the AI behind the Goldnat AI Assistant for XenForo — a major provider's API, a hosted open model, or your own server — with cost, quality, privacy and effort for each, and the open models that passed our test. Provided as a service, without warranty.

📅 Updated: 2026-11-17

The Goldnat AI Assistant for XenForo does not include an AI model. You connect it to one. This page helps you choose which, and shows what each choice costs.

Please read this first. This page is offered as a service only, without any warranty. Prices, models and licences change. Results depend on your forum, your hardware and your settings. Before you decide, check the current prices and the licence terms with the provider yourself.

Three ways to run it

1. API of a major provider 2. Open model, hosted, through an API 3. Your own server
Who it suits A forum that wants the simplest setup. A good place to start. Forums that want the lowest bill, or an open model, without running a server. Forums whose content must not leave their own hardware, or very busy forums.
Cost a month — smallabout 120 commands a month $0.04 to $0.16 $0.01 to $0.04 from about $250, at any sizeplus a one-time setup fee of about $245
Cost a month — mediumabout 1,200 commands a month $0.40 to $1.63 $0.14 to $0.44 from about $250, at any size
Cost a month — largeabout 12,000 commands a month $3.98 to $16.26 $1.41 to $4.41 from about $250, at any size
Answer quality High. These models were not part of our comparison. High with the right model: English 92, other languages 92 to 95 out of 100 in our test. The same models as in way 2. A compressed copy on your server can be slightly weaker.
Does forum content leave your server? Yes, to the provider. Yes, to the hosting service. No.
Effort A few minutes: create a key, paste it. A few minutes: create a key, paste it, choose the model. A server with a graphics card to rent, set up, update and secure.

What small, medium and large mean: the number of AI commands members run in a month — about 120, 1,200 and 12,000. We arrived at these from 200, 2,000 and 20,000 active members a month, one in ten of them using the Assistant 6 times; count your own commands instead if you know them. Costs are estimates for an English-language forum with short threads. Way 1 is the range of the low-cost models of OpenAI, Anthropic and Google; way 2 the range of the open models we recommend. Prices read 2026-11-17. The server price is the vendor's published price (Hetzner GEX45 dedicated server, one 24 GB graphics card: EUR 214 a month plus a one-time EUR 209 setup fee, excluding VAT), shown in dollars; the setup fee is paid once and is not in the monthly figure.

Our advice: start with way 1 or way 2. Both cost little, need no server, and you can switch later in a few minutes. Choose your own server only if forum content must stay on your hardware, or once your monthly bill is close to the price of a server.

Way 1 — the API of a major provider

You create an API key at OpenAI, Anthropic or Google and paste it into the Assistant. The Assistant suggests a model for you; a low-cost one is a good place to start.

  • Good for: a forum that wants to be running in a few minutes and is content with a well-known provider.
  • Keep in mind: posts are sent to the provider to be processed. Read the provider’s data terms if that matters to you.

Way 2 — an open model, hosted for you

Open models are free to use, and hosting services run them for you and charge per use, like the big providers. You get one key from the service and choose the model. We tested this way with OpenRouter.

Two open models did as well as anything we tested, in English and in other languages: Gemma 4 (31B) and Mistral Small 3.2 (24B). They are the first two rows of the table in the next section.

  • Good for: the lowest bill without a server, or a preference for open models.
  • Keep in mind: posts are sent to the hosting service, and the service decides which data centre runs the model.
  • A cheaper hosted model for English-only forums: GLM-5.3 Flash from Z.ai. It did well in English but clearly worse in other languages, and some runs returned no answer. We do not recommend it for anything else.

Way 3 — your own server

Here nothing leaves your hardware. In return you rent and look after a server with a graphics card, and you pay for it every month whether members use the Assistant or not.

We tested 21 open models and list only those that passed. The full results are at the end of this page.

Server Model (Ollama name) Memory it needs English Other languages Runs finished Time per command Licence
Mid-size serverOne graphics card with 24 GB of memory gemma4:31b about 22 GB 97 95 72 of 72 3 s Apache 2.0
Mid-size serverOne graphics card with 24 GB of memory mistral-small3.2:24b about 17 GB 97 92 72 of 72 4 s Apache 2.0
Mid-size serverOne graphics card with 24 GB of memory gemma4:26b about 21 GB 92 95 71 of 72 4 s Apache 2.0
Small serverA graphics card with 12 to 16 GB of memory gemma3:12b about 10 GB 93 82 67 of 72 9 s Gemma Terms of Use (custom) — check the terms
Smallest server we can recommendA graphics card with 12 GB of memory qwen3.5:9bwith thinking off about 10 GB 86 70 22 of 24 4 s Apache 2.0
Large serverOne graphics card with 96 GB of memory, or two with 48 GB mistral-medium-3.5:128b about 82 GB 96 97 72 of 72 3 s Modified MIT — check the terms
Large serverOne graphics card with 96 GB of memory, or two with 48 GB gpt-oss:120b about 67 GB 90 90 71 of 72 5 s Apache 2.0

English, other languages (0 to 100), runs finished and time per command are what we measured, on a hosting service that runs each model at full precision. Memory and licence are published by the model's maker, not measured: memory is the size of the copy Ollama downloads, plus about 2 GB of working room.

Room for long threads (published figures, not measured): a model also needs memory for the text it is reading. gemma4:31b is listed at 19-20 GB, which leaves little spare room on a 24 GB card. If your forum has long threads, choose mistral-small3.2:24b (15 GB) on that card, or a card with more memory.

How to choose: take the largest row your server fits. The mid-size models are the sweet spot — in our test they scored as well as the large ones, so a larger server buys very little for this Assistant. Below the smallest row we found no model we can recommend.

For some languages a language-specialised open model exists, for example DictaLM for Hebrew. We could not measure such models in this test.

What kind of server

The model runs on a graphics card. What matters is the card’s memory: the model must fit in it.

ServerWhat it runsPublished priceSource
One graphics card with 24 GB of memory, rented by the monthThe mid-size and smaller modelsHetzner GEX45: EUR 214 a month, plus a one-time setup fee of EUR 209, excluding VATvendor's page, checked 2026-11-17
The same kind of card, rented by the hourThe mid-size and smaller modelsRunPod, RTX 4090: USD 0.34 an hour (Community Cloud) to USD 0.89 an hour (Secure Cloud), no setup fee listed; about USD 250 to 650 a month if it is always onvendor's page, checked 2026-11-17
One graphics card with 96 GB of memory, rented by the monthThe large modelsHetzner GEX63: EUR 999 a month, plus a one-time setup fee of EUR 999, excluding VATvendor's page, checked 2026-11-17

All three prices are published by the vendors and were checked on their own product pages on the date shown. We did not rent these servers. The monthly server figure used on this page (about $250) is the first row in dollars and does not include its one-time setup fee (about $245); over the first year the two together come to about $270 a month.

A server without a graphics card can run the two smaller models if it has 16 GB of memory, but slowly: an answer can take a minute or more, and the site on the same server slows down while the model works. Shared hosting cannot run a model at all.

Connect your own server in five steps

  1. Get the server from the table above, sized for the model you chose.
  2. Install a model server and download the model. We describe Ollama: follow its guide for Linux or Docker, then download the model by the name in our table from the Ollama model library.
  3. Keep it private. Ollama listens only on the server itself (127.0.0.1:11434) and has no login. Leave it that way. If the model runs on a different machine than your site, put it behind an HTTPS address that requires a token.
  4. Enter it in the Assistant. In the admin panel open AI Assistant › AI provider. Choose your own server, enter its address (http://127.0.0.1:11434/v1 when it is on the same machine as the forum) and the model name. Leave Model thinking on Off.
  5. Press Connect and test. The Assistant sends a real request and tells you in plain words if something is wrong.

This is your own setup. We do not install, run or support the model server; LM Studio and vLLM work the same way.

The “Model thinking” setting

Some models “think” before they answer — 13 of the 21 open models we tested do. The thinking counts against the length of the answer, so the answer can come back empty and the command fails. In our test this was the most common reason a model failed.

From version 1.9.1 the provider page has a Model thinking setting. It is Off unless you change it, and that is the safe choice. One model in our table is recommended only with thinking off; with it on, it failed more than a quarter of the runs.

A few models cannot stop thinking. If Connect and test reports an empty answer while the setting is Off, choose another model from the table.

What will it cost me?

On this page small, medium and large mean the number of AI commands members run in a month: about 120, 1,200 and 12,000. One command is one click on an Assistant button: a summary, a question, a translation and so on.

With an API (ways 1 and 2) you pay for the text that goes in and out of the model. The Assistant adds nothing on top. With your own server you pay for the server.

  • Text per command — measured. We recorded how much text each of the eight commands uses on our test forum. Real threads vary; a long thread costs more to summarise.
  • Prices — official price lists, read on 17 November 2026: OpenAI, Anthropic, Google and OpenRouter.

Cost per 100 runs of each command

Estimated cost in US dollars of 100 runs of each command (English forum, short threads). Prices read 2026-11-17.
Model Summarize threadAsk the forumCatch-upUnanswered questionsTrendingTranslate (5 short posts)Duplicate checkRephrase
OpenAI
GPT-6 Luna $0.06$0.03$0.03$0.02$0.02$0.04$0.03$0.01
GPT-5 mini * $0.23$0.12$0.10$0.07$0.08$0.13$0.09$0.04
Anthropic
Claude Haiku 5.5 $0.06$0.03$0.03$0.02$0.02$0.04$0.03$0.01
Claude Sonnet 5.5 $1.30$0.67$0.55$0.41$0.48$0.72$0.54$0.23
Google
Gemini 2.5 Flash-Lite $0.06$0.03$0.02$0.02$0.02$0.03$0.02$0.01
Gemini 3.1 Flash-Lite $0.18$0.09$0.08$0.06$0.07$0.10$0.08$0.03
Open and low-cost models, hosted (through OpenRouter)
Gemma 3 12B (open model, hosted) $0.02$0.007$0.006$0.008$0.008$0.008$0.007$0.005
Mistral Small 3.2 24B (open model, hosted) $0.04$0.01$0.01$0.02$0.02$0.02$0.02$0.008
Gemma 4 31B (open model, hosted) * $0.07$0.02$0.02$0.02$0.03$0.03$0.03$0.01
GLM-5.3 Flash (Z.ai) * $0.02$0.007$0.009$0.007$0.02$0.007$0.010$0.004

* A model that thinks before it answers. The thinking is billed too, so the real cost can be higher than shown.

Cost simulator

An estimate, not a quote. Everything is calculated in your browser; nothing you enter is sent anywhere. The size of a forum here is the number of AI commands members run in a month; the member fields below are one way to arrive at it.

Your forum
Model
Commands a month, by command (edit to match your forum)

AI commands a month—

Estimated cost a month — likely between — and —

About — per 100 commands.

A server of your own that can run a comparable open model: from about $250 a month, plus a one-time setup fee of about $245 — about $270 a month over the first year (Hetzner GEX45 dedicated server, one 24 GB graphics card: EUR 214 a month plus a one-time EUR 209 setup fee, excluding VAT; published by the vendor, read 2026-11-17; fits open models up to about 30B parameters).

When does your own server pay off?

A server costs the same every month. An API costs more as members run more commands. The simulator above shows where the two meet for your number of commands. It shows two figures: with the monthly price alone, and with the one-time setup fee spread over the first year.

Why our whole test cost about two dollars

Our whole test — 1,896 commands on 21 open models and two hosted ones — cost us $2.12. A forum's monthly bill can be much higher or much lower, because the bill is simply the number of commands times the price of one command. With the open models we recommend, hosted, one command costs about 0.01 to 0.04 of a cent in our estimate. With a premium model (Claude Sonnet 5.5) it costs about 0.8 of a cent. So 12,000 commands a month come to roughly $1.41 to $4.41 on an open hosted model, and roughly $92.28 on the premium one.

An AI cost page in the admin panel is being added to the Assistant. It shows what you really spent this month and tells you when a server of your own would cost less. If your version has that page, trust it over any estimate here.

Read the cost numbers as estimates

  • Prices change. Check the provider’s price page before you decide.
  • Repeat views are free. The Assistant keeps a thread summary and shows it again without calling the model until enough new posts arrive.
  • Models that think cost more. The thinking is billed as output. The table marks such models with an asterisk.
  • Languages that are not written in Latin letters use more tokens for the same amount of text. Choose the forum language in the simulator.

How we tested, and the limits

  • We ran the Assistant’s eight commands (thread summary, ask the forum, catch-up, unanswered questions, trending, translation, duplicate check, rephrase) on a made-up test forum: twelve cases in English and twelve in one other language, three rounds, 72 runs per model. Assistant version 1.8.0, 17 November 2026.
  • Scores are 0 to 100. They combine: did the command finish, was the answer complete, were the cited posts real, was the format valid, was the language right. In one of the three rounds a strong reviewer model also marked every answer.
  • “Other languages” is one language. A model that did well there will not necessarily do as well in yours. Test with your own forum before you commit.
  • We measured full-precision copies run by a hosting service. The copy a model server downloads by default is compressed to fit in less memory and can be somewhat weaker.
  • Time per command is the hosting service’s speed, not your server’s.
  • It is a small test. A few points between two models mean nothing. A gap of 15 points or more does.
  • Check the licence yourself before you use an open model on a business site. Licences change, and some models that look free are released for non-commercial use only.

Once more: no warranty

Everything on this page is offered as a service only, without any warranty. It describes one test on one date. Prices, models and licences change, and your results depend on your forum, hardware and settings. Check current prices and licence terms with the provider before you decide.

Full results of the test (21 open models)

72 runs per model: eight commands, twelve cases in English and twelve in another language, 3 rounds. Assistant 1.8.0, measured on 2026-11-17. Scores are 0 to 100.

ModelScoreEnglishOther languagesRuns finishedFormat errorsInvented sourcesTime per command
gemma4:31b 969795 72 of 7200 3 s
mistral-medium-3.5:128b 969697 72 of 7200 3 s
mistral-small3.2:24b 959792 72 of 7230 4 s
gemma4:26b 939295 71 of 7200 4 s
gpt-oss:120b 909090 71 of 72150 5 s
gemma3:12b 879382 67 of 72140 9 s
qwen3:14b 848980 70 of 72310 12 s
phi4:14b 828084 67 of 7217 5 s
llama3.3:70b 788076 62 of 72260 7 s
qwen3.5:122b 778766 62 of 72226 35 s
gpt-oss:20b 757673 64 of 72360 7 s
mistral-nemo:12b 707861 67 of 7212 9 s
llama3.1:8b 687463 66 of 72152 15 s
command-r7b:7b † 687858 58 of 72211 4 s
qwen3.5:9b 667557 52 of 72360 11 s
qwen3.8:27b 637353 51 of 72470 8 s
muse-glimmer:30b 606653 50 of 72670 21 s
llama3.2:3b 476034 46 of 72490 5 s
qwen3.6:35b 364229 32 of 72740 20 s
granite4.2:8b 252921 25 of 72780 28 s
nemotron-3.5-lightning:30b 222518 21 of 72820 10 s

† Released for non-commercial use only. Tested, but not a candidate for a business site.

The same model with thinking turned off

One round each. Where a model cannot stop thinking, it was set to short thinking instead.

ModelScore as it comesRuns finishedSettingScore with the settingRuns finished
granite4.2:8b 2525 of 72 thinking off 6421 of 24
qwen3.5:9b 6652 of 72 thinking off 7822 of 24
qwen3:14b 8470 of 72 thinking off 8124 of 24
gpt-oss:20b 7564 of 72 short thinking 8123 of 24
qwen3.8:27b 6351 of 72 thinking off 9124 of 24
qwen3.6:35b 3632 of 72 thinking off 8323 of 24
muse-glimmer:30b 6050 of 72 short thinking 8021 of 24
nemotron-3.5-lightning:30b 2221 of 72 thinking off 6921 of 24
gpt-oss:120b 9071 of 72 short thinking 8624 of 24
qwen3.5:122b 7762 of 72 thinking off 9024 of 24

Hosted low-cost models

ModelScoreEnglishOther languagesRuns finishedTime per command
glm-5.3-flash 799465 61 of 728 s
glm-5.3 707861 54 of 729 s

Not measured, because the hosting service we used does not offer them: phi4-mini:3.8b, granite4.2:3b, qwen3.5:4b, gemma4:e4b, gemma4:12b, dicta-il/DictaLM-3.0-24B-Thinking.

This page is informational and non-binding. It describes the general design of the AI Connect plugins; actual behavior can vary between platforms and plugin versions. For the binding terms, see the Terms & Conditions.