Which LLM service is right for you?
The University offers several ways to use large language models — from a ready-to-use chat assistant to programmable gateways and HPC batch pipelines. This guide explains the basics and helps you pick the offering that fits your work.
What is a large language model?
A large language model (LLM) is an AI system trained on very large amounts of text so it can understand and generate human-like language. You give it a prompt — a question, a document, some code — and it responds with text: an answer, a summary, a draft, a translation, or an explanation. Popular examples include GPT, Claude, Gemini, and open-weight models such as Llama and Qwen.
You typically use an LLM in one of two ways: through a chat interface in your browser, or through an API that lets your own programs and scripts send prompts automatically.
Why they’re useful
- Save time on writing and editing — draft emails, summarize long documents, rephrase for clarity.
- Understand and write code — explain unfamiliar code, generate scripts, help debug.
- Answer questions over your own material — some services can search course notes, papers, or documents and answer with citations (retrieval-augmented generation, or RAG).
- Process data at scale — classify, extract, or transform thousands of records automatically through an API or batch pipeline.
- Support teaching and research — tutoring assistants, literature triage, hypothesis brainstorming.
Be careful with your data
Where your data goes matters
When you send text to an LLM, that text leaves your computer and is processed on a server somewhere. If that server is operated by a vendor outside the University, your prompt — and anything you paste into it — is disclosed to that third party and may be logged, retained, or used to improve their models. Before you paste anything, ask: where is this being hosted?
- Never paste regulated or sensitive data into a non-Illinois-hosted service unless a University agreement explicitly covers it — FERPA (student records), HIPAA (health data), export-controlled research, unpublished IP, security credentials, or personally identifiable information.
- Prefer Illinois- or NCSA-hosted options for sensitive work. Several offerings below run entirely on University or NCSA infrastructure, so your data stays within institutional control.
- Treat commercial models routed through a gateway as external. Even when reached through an Illinois service, a prompt sent to OpenAI, Anthropic, or Google is still processed by that vendor.
- Verify important output. LLMs can produce fluent but incorrect information (“hallucinations”). Check facts, code, and citations before you rely on them.
- When in doubt, consult the Office of the CIO and your unit’s data-classification guidance.
The four offerings
Each card links to a detail page and to the service itself. All four are built or operated at the University of Illinois / NCSA.
IllinoisChat
Chat & Q&A over your own material
Best for: instructors and students who want a course assistant.
Lumen
Illinois-hosted models, ready to use from chat or code
Best for: individuals and teams that want to just use provisioned models.
LLM Hub
On-demand model hosting on NCSA HPC
Best for: teams that want private, self-hosted models with a dashboard.
LLMFlux
Batch LLM processing on HPC clusters
Best for: large offline datasets processed on Illinois HPC.
Still deciding?
Answer five quick questions and we’ll recommend the service that best fits your work.
Frequently asked questions
Which service should I use for sensitive or regulated data?
LLM Hub, LLMFlux, and Lumen all run models entirely on Illinois/NCSA infrastructure, so your prompts and data stay within institutional control. IllinoisChat is safe only when pointed at an institution-hosted (or in-browser) model, because it can also route to commercial vendors. Never paste FERPA, HIPAA, export-controlled, or other regulated data into a non-Illinois-hosted service without an approved agreement.
Can I use commercial models like GPT, Claude, or Gemini?
Yes, through IllinoisChat, which can route to OpenAI, Anthropic, Google, and others. Remember that when you choose a commercial model your prompt is sent to that outside vendor and may be logged or retained. The other services — including Lumen — run only Illinois/NCSA-hosted models, so if you need commercial models, IllinoisChat is the option.
Do I need to know how to code?
Not for IllinoisChat, which is a web app for asking questions over your own documents, or for the chat interfaces built into LLM Hub and Lumen. You do need to be comfortable on the command line and with HPC/SLURM to use LLMFlux or the programmatic APIs.
What does it cost, and are there spending limits?
All four are free to use — they are built or operated at Illinois/NCSA and, like other Illinois Computes resources, carry no charge; there is no per-token billing. Only Lumen enforces usage limits, metering per-user and per-group usage in virtual “coins” (not dollars). LLM Hub, LLMFlux, and IllinoisChat track usage but do not cap it, so demand is managed through HPC/SLURM allocations and deployment limits.
How do I get access?
Most services use Illinois federated login (CILogon or Keycloak SSO), so your NetID gets you in. LLMFlux instead relies on your existing HPC/SLURM account and filesystem permissions. Each service’s detail page links straight to its sign-in or documentation under “Get started.”
What’s the difference between a chat interface and an API?
A chat interface is a website where you type prompts and read replies — no setup beyond logging in. An API is an endpoint your own programs and scripts call to send prompts automatically, which is how you process data at scale or build the model into an application. LLM Hub and Lumen offer both; IllinoisChat is primarily a chat app; LLMFlux is API/command-line only.
Is there a technical comparison of the underlying software?
Yes. These offerings are built on different software packages, and we keep a detailed, feature-by-feature comparison of that software — models, authentication, budgets, observability, deployment, protocols and more — including a few related tools beyond these four offerings. See the software comparison table.