Choosing and connecting an LLM

The intelligence in bAIbel AV — the translation, the alignment, the placing of codes and tags — comes from a large language model, or LLM. bAIbel AV does not include one. You connect the model of your choice, and that choice is yours to make per project. Models come in many shapes and sizes, run by very different kinds of organisation, with very different privacy implications. This guide explains the landscape in plain terms, so you can match the model to the job in front of you.

You are not exposed by default

Remember the privacy stage from guide 1. bAIbel AV can hide confidential names and numbers before any text reaches a model. That protection applies whichever model you choose. Choosing your provider well is the second layer, not the only one.

Two questions for every project

Choosing a model comes down to two simple questions.

Question	What it decides
Who runs the model?	Your privacy. It decides whose servers your text passes through, and under whose rules.
How big a model?	Your quality and cost. Bigger models are more capable and cost more; smaller ones are cheaper and often enough.

The rest of this guide answers the first question with three groups of provider, then returns to the second.

Three groups of provider — a travel metaphor

Think of getting your text to a model as a journey it has to take. There are three ways to make that journey, and they trade convenience, power, and control differently.

Figure 1. Three ways to reach a model. Fly with a flagship airline, book through a travel hub, or take your own vehicle. Each trades convenience, power, and control differently.

Group 1 — The big labs (fly with a flagship airline)

These are the companies that build the best-known models: OpenAI, Google Gemini, and Anthropic Claude. They offer a full range, from enormously powerful flagship models down to tiny, fast “nano” models for everyday tasks. Like a flagship airline, they are powerful and go almost anywhere — but a very large company is handling your luggage, under its own terms.

Their data policies depend on the plan you are on, and they change. Free consumer apps often use your conversations to improve future models. Paid programming interfaces and business plans usually do not train on your text by default, though they may keep it for a short time to detect misuse, and may have staff review content that gets flagged. The lesson is not “avoid the big labs”; it is “check the current policy for the exact plan you use”.

Group 2 — The platform providers (book through a travel hub)

These companies do not usually build their own models. Instead they host many models — open-source and closed-source alike — and charge you to use them. Fireworks.ai, Together.ai, and OpenRouter are well-known examples. Like a travel hub that books you onto dozens of airlines with one membership, a single account gives you instant access to a huge range of models.

Two things to keep in mind. First, their catalogue changes often — models appear and disappear, so a model you used last month may move. Second, on privacy you must check two things: the platform’s own policy, and, for closed models, the policy of the company that made the model. Several platforms hold a recognised security audit (see SOC 2 below), but you must confirm it — and the free, community models offered on a service like OpenRouter are not suitable for confidential work.

Group 3 — Run it yourself (take your own vehicle)

The third group keeps the journey under your control. You run an open model on hardware you command: on your own machine with software such as Ollama, where nothing leaves the computer; or on a powerful GPU you rent by the second from a service such as RunPod, which holds a recognised security audit. Like driving your own car or renting one, you decide the route and who is in the vehicle.

The trade-off is effort and ceiling. You handle the setup, and the open models you can run yourself may not quite match the very best flagship models from the big labs. For the most confidential work, that trade is often well worth making.

The three groups compared

Group	Model range	Privacy posture	Always verify	Best for
Big labs OpenAI, Gemini, Claude	The most capable models, flagship down to nano.	Your text reaches a large AI company. Consumer tiers may train on it; paid API and business plans usually do not by default.	The data-use and retention policy of your exact plan.	Highest quality; non-confidential or well-masked work.
Platform providers Fireworks, Together, OpenRouter	A vast, changing catalogue of open and closed models.	Your text reaches the host, and for closed models the model’s maker too. Often security-audited.	The platform’s audit status; never use free community models for sensitive data.	Range and flexibility; trying many models from one account.
Run it yourself Ollama, RunPod	Open models you host; ceiling below the top flagships.	The most control. With Ollama, text never leaves your machine; with RunPod, it stays in a GPU you rent and control.	Your own setup; for RunPod, its audit and region.	The most confidential work; full control over data.

The second question: how big a model?

It is tempting to reach for the largest, most famous model every time. For translation, you rarely need to. Much of the everyday work — placing codes and tags, aligning segments, smaller edits — runs perfectly well on small, cheap models. Save the big flagship models for the passages that genuinely need their extra judgement.

It costs less than you expect

You pay only for what you use, and a little goes a long way. Setting up an account with a provider and adding five to ten dollars of credit will usually cover a large amount of translation work. You can start small and top up only if you need to.

What “SOC 2” means

You will see the term SOC 2 when comparing providers. In plain terms, it means an independent auditor has examined how a company protects the data it handles and confirmed it meets a recognised standard. It is a useful signal that a provider takes security seriously.

Two cautions. A provider being SOC 2 audited covers that provider’s service, not necessarily every model it resells. And free or community offerings are usually outside the audit altogether. Treat SOC 2 as a box to confirm for the specific provider and plan you intend to use — not as a label you can assume.

Matching the choice to the project

The right choice depends entirely on the situation, and two real projects can call for opposite answers.

A travel guide, already published in its source language

The source text is already public. There is little confidential about it. Here you can optimise freely for quality, cost, and convenience. A capable flagship model from a big lab, or a strong model through a platform provider, is a fine choice, and you need not agonise over privacy.

A company’s confidential earnings release, before publication

This is the opposite case. The text is market-sensitive and must not leak. Here you combine both layers of protection. Turn on bAIbel AV’s privacy stage so names and figures are masked before anything is sent. Then choose a provider to match: a security-audited platform with clear no-training terms, or — for the strongest guarantee — run the model yourself with Ollama or on a rented RunPod GPU, so the text stays under your control. Avoid free models and consumer chat tiers entirely.

Figure 2. Match the provider to the document. Public, low-risk text can use the most convenient and capable option; confidential text calls for masking plus a self-hosted or audited provider.

There is no single right answer

Every project can call for a different solution. bAIbel AV is built for this: you can set a different provider for each project, and even a different model for different tasks within a project.

Setting this up in bAIbel AV

Open LLM Configuration from the sidebar. The screen, API keys and service configuration, is organised into the same groups described in this guide:

Active foundation model providers — the big labs (group 1).
Active platform providers — the hosting platforms (group 2).
Active GPU rental platforms and Active LLM model software platforms — running it yourself (group 3).

You connect a provider by adding its key in Configuration and key files. Once connected, a provider moves up into its “Active” group.

Figure 3. The LLM Configuration screen. Providers are grouped exactly as in this guide, so the three groups map straight onto what you see.

To choose which model does the work, use Default Provider for everything, or Role-Specific Providers to assign different models to different tasks — for example a small model for placing tags and a flagship for translation. Picking a model is a two-step choice: Choose an LLM Provider, then Choose a Model. Models you have not added a key for are marked so you can see them but cannot select them.

Figure 4. Role-Specific Providers. Assign a different model to each task, then use Change… to pick the provider and model.

Terminology used in this guide

LLM (large language model): The AI model that provides the intelligence for translation, alignment, and placing codes. bAIbel AV connects to one of your choosing.
Foundation model provider: A company that builds and runs its own models, such as OpenAI, Google Gemini, or Anthropic Claude.
Platform provider: A company that hosts many models, open and closed, and charges for access — such as Fireworks, Together, or OpenRouter.
Self-hosting: Running a model on hardware you control, with software such as Ollama on your own machine or a rented RunPod GPU.
SOC 2: An independent audit confirming a provider meets a recognised standard for protecting data. Verify it per provider and plan.
Role-Specific Providers: The setting that lets you assign different models to different tasks within a project.