EleosLocal AI on Windows, offline and on your own machine
Eleos downloads an open model, runs it on your own graphics card or processor, and puts it to work on your files. No AI subscription to buy, and the conversation never leaves the PC.
What you get
- No AI subscription. The model is open and costs nothing to run; you pay for Eleos, not for tokens.
- Offline. After the download, it works with the network unplugged.
- No terminal. Pick a model from the list; Eleos fetches it, checks it, and starts it.
- The same agent. Local chats read, write and organise files like the cloud ones do.
What is in the list
One model today: Qwen3.5 2B, in four quality levels from 1.3 GB to 2.0 GB, so a smaller file fits a smaller graphics card. The list is signed and served by Eleos; the model runs through llama.cpp.
One, rather than a menu, because of the floor below: a model has to hold about 28,700 tokens before an agent can use it at all, and most small open models are published with far less. Adding one that cannot answer is not a choice, it is a disappointment with a download attached.
What a local model actually needs
The download is the small part. What decides whether a model answers at all is how much context it has to hold: an agent sends its tools and your files with every turn, so the window matters more than the parameter count.
Three things we had to fix before local models answered reliably, all measured rather than guessed:
- Thinking has to be off for small models. A 2B model asked "17 times 23" spent 5,458 hidden tokens thinking and never reached an answer. With thinking disabled the same request answered.
- The context floor is real. The first request from the agent harness is about 25,200 tokens with every tool available, and about 6,144 with the local set. So Eleos will not start a model published under 28,672 tokens: below that the first message is refused on any machine, however much memory it has.
- The window the CLI believes. For a model name it does not recognise, the maximum-context setting is the window, and at 32,768 it compacted before the first message every time.
Full write-up: ten dead ends in the first five minutes.
What to expect
Small local models are slower and less capable than the large cloud ones. They suit private documents, simple file work, and anything that must stay on the machine. When a job needs more, the same chat can switch to a cloud plan you already pay for.
Questions
What does it cost?
The model is open and costs nothing to run; Eleos is $19.90 a month after seven days free, with no card to start. What you save is the AI subscription: no Claude or ChatGPT plan is involved when a model runs on your own PC.
Do I need a graphics card?
It helps. Models run on the processor too, slower. About 8 GB of memory is the comfortable floor for the smaller models; Eleos says which ones fit your machine before you download. What each build actually costs to load, measured: what fits in 8 GB of VRAM.
Does anything leave my PC?
Your conversation does not. The model runs on your machine and the text never leaves it. Eleos itself talks to our servers for two things — checking your subscription is live, and counting that the app started — and neither carries any part of what you typed. The count is the release number and whether this machine had started it before, with no identifier attached.
Which quality level should I start with?
The smallest, Q4. It downloads fastest and answers on the widest range of machines; move up only if the answers are not good enough.
Is this the same as LM Studio or Ollama?
Those run a model and let you chat with it. Eleos runs the same kind of model and gives it a job: reading and writing your files, using apps when you allow it, on a schedule if you want one.
Where to go next
Try it on Windows. Seven days free, no card, then $19.90/month for two PCs. The AI is whichever plan you already pay for, or an open model on your own PC, which needs no subscription.