01
The model is the one part you should be free to change
The best AI model this year will not be the best one next year. Prices move, terms change, and a provider can retire the very model you built on.
That would not matter much if a model were easy to replace. In most AI products it is not. The prompts, the tool definitions and the safety rules are all written for one vendor's model, so a change of model is a rebuild.
Smart Forge is built the other way round. Your tools, agents, guardrails and skills are defined once, outside any model, and stored as files you own. The model is plugged in underneath them. You can run your own, use the one your cloud contract covers, or connect a provider directly, and you can change your mind later without touching anything you built.
Defined once, used with any model
02
Three places a model can run
Where the model runs decides where your prompts and your data go. There are three choices, and they suit different situations. Pick one below to see where the prompts end up.
Route 1On your own servers
An open-weight model, such as Qwen, Gemma, Llama, Nemotron, Mistral or DeepSeek, runs on hardware you control. Smart Forge needs only the address of that server. Nothing is sent to anyone outside. This is the route when the data must stay inside, or when you already run the infrastructure for it.
Route 2Inside your cloud account
If your company has already approved AWS or Azure, models are available there through services such as AWS Bedrock or Azure AI Studio. Prompts go to your own cloud account, under the contract, the billing and the security controls your cloud team already manages. There is no new vendor to approve.
Route 3With a hosted provider
You connect OpenAI, Claude or Gemini with your own key. It is the quickest way to start, and it gives you the largest models, which are the strongest at long investigations with many tool calls. An enterprise agreement with the provider can state that your data is not used to train its models.
03
Six reasons to run your own model
Hosting a model yourself is more work than using a provider, so it has to earn its place. These are the reasons that usually decide it.
Prompts, tool results and answers stay inside your network. For health records, proprietary designs and other regulated data, that can be the difference between using AI and not being allowed to.
In Reasoning Mode the model reads your results in order to think about them. With a model inside your walls, it can read data you could never send to a provider. About Reasoning Mode.
If you already run GPU servers, or have a cloud contract in place, the security review is behind you. No new vendor has to be assessed before the first question is asked.
A provider can update or retire a hosted model. A model you host stays the same version until you choose to move, so answers do not shift under you.
A hosted model is billed for every token it reads and writes. A model you host costs the hardware it runs on, however much it is used, which suits steady, high-volume work.
A model on your own servers processes data where those servers stand. Where the rules say data may not leave a country or region, that settles the question.
04
Adding a model takes minutes, not a project
There is no migration and nothing to rebuild. You tell Smart Forge where the model is, and then you choose it.
Step through it below.
The choice is made for each chat and for each scheduled run, not once for the whole company. One team can use a hosted model while another keeps everything inside, and the same person can use a small model for a quick look-up and a large one for an investigation.
05
Change the model. Keep everything you built
This is the part that makes the rest possible. A tool in Smart Forge is a definition and a piece of code. An agent is a set of settings. Guardrails are settings, and skills are documents. None of them is written in a way that only one model understands.
So when you change the model, they carry over as they are. The same tool is called in the same way and the same guardrails apply. The figures in your reports do not change either, because they are computed by queries on your data and not by the model.
The model is a choice you can revisit. What you built is yours to keep.
The usual way, and the Smart Forge way
The openness runs in the other direction too. Your tools are served over an open standard, so other AI assistants your people already use can call them. Read about the development ecosystem.
06
What to weigh before you choose
Each route has a cost, and it is better to choose with all of them in view.
| Compared on | Your own serversSelf-hosted | Your cloud accountAWS Bedrock, Azure AI Studio | A hosted providerOpenAI, Claude, Gemini |
|---|---|---|---|
| Where prompts go | Nowhere. They stay in your network. | To your own cloud account | To the provider |
| Who runs the model | You do | Your cloud provider | The model provider |
| Strongest at | Sensitive data, and steady high-volume work | Using a contract and controls you already have | The hardest investigations, and a quick start |
| What to weigh | You run the hardware. Open models are smaller, and weaker at complex tool use. | The models on offer depend on your cloud service. | Data leaves your network, under your agreement. |
You do not have to pick one. A sensible arrangement is a model of your own for anything sensitive and a hosted one for the rest. And where no data at all should reach a model, there is Operational Mode.
Self-hosted models are in beta. Your compliance team decides what applies to your data.
07
Size the model to the work
With models, you get what you pay for, in hardware or in tokens. The good news is that most work does not need the largest one.
Everyday questions ask a model to do something fairly narrow: understand the question, pick the right tool and fill in its arguments. A mid-sized open-weight model, around 30 billion parameters, such as one from the Qwen family, handles this well, and it runs on hardware many companies can host themselves.
Detailed analysis is different. When the model has to plan several steps, read results from different systems, notice what does not add up and explain it, a larger model gives clearly better answers. That is where a bigger self-hosted model, or a top model from a provider, earns its cost.
Look-ups, lists, status questions and Operational Mode. A model around 30B, self-hosted, is usually enough.
Reasoning Mode, multi-step investigations and teams of agents. A larger model pays for itself here.
Because the model can be chosen per chat and per scheduled run, you do not have to pick one size for everything. And you do not have to guess. We look at the questions your teams ask and the data involved, and recommend the models that fit, from what you can host to what you would rent. Talk to us about your needs.