The part every other tool leaves you to find out the hard way.
51 curated models · sized against your actual RAM · pulled with one click
The usual way to discover a model is too big for your machine is to download it and watch everything freeze. adris reads your available memory and marks each model as comfortable, tight or out of reach before you commit to the download — so the first local model you try is one that runs.
The catalogue spans roughly 0.7 GB models that run on a modest laptop up to frontier-scale open weights, including Llama, Mistral, Phi, Gemma and Indic models tuned for Hindi, Tamil and other Indian languages.
No key, no account, no request out.
A local model means the weights sit on your disk and inference happens on your processor. There is no API key because there is no API — nothing is being called. That is the difference between "we don't train on your data" (a policy) and "your data never left the building" (a fact about the network).
It also means no per-token cost. Once the model is on your machine, running it is free for as long as you use it, which changes what is worth automating.
Ask two models the same thing, side by side.
Model choice is hard to reason about from benchmarks. adris can put two to four models on the same prompt at once and show the answers together with timings, so the comparison is on your work rather than someone else's leaderboard.
If a model is too large for one machine, Mesh pools memory across several computers you own so you can run it anyway. See how Mesh works →
A local model that can actually do things.
A chat window is the least interesting thing to do with a local model. In adris the same model drives the agents, the automation steps, the code assistant and the document reading — so the private model is not a sandbox off to one side, it is the engine for the whole application.
Against Ollama, LM Studio and the cloud.
| adris | Ollama | LM Studio | ChatGPT / Claude | |
|---|---|---|---|---|
| Runs offline | Yes | Yes | Yes | No |
| Command line needed | No | Yes | No | No |
| Tells you what your PC can run | Yes, before you download | No | Partly | n/a |
| Cost per use | Free once downloaded | Free | Free | Per token or per seat |
| Pool RAM across machines | Yes (Mesh) | No | No | n/a |
| The model can also do work | Agents, automations, docs | Chat + API | Chat + API | Chat + tools |
| Frontier-level quality | Included with your plan, or your own key | No | No | Yes |
Ollama and LM Studio are good at what they do and this is not an attempt to talk you out of them. The difference is what happens after the model is running: here it drives the agents, the automations and the document reading rather than sitting in a chat window on its own.
Common questions
Do I need a GPU?
No. A GPU makes local inference faster but is not required. The catalogue marks which models are realistic on CPU.
Does it use Ollama?
It can run models through Ollama if you have it, and it manages the pull and the runtime for you if you do not.
How much RAM do I need?
8 GB is enough for small models, 16 GB is comfortable for mid-size ones, and Mesh lets you pool memory across your own machines for anything larger.
Is it really free?
The Models module is free forever, including on the free tier. Running a model on your own hardware costs nothing per use.
Can it work with no internet at all?
Yes, once the model is downloaded. Inference is entirely local after that.
Which model should I start with?
Whichever the catalogue marks comfortable for your machine. A small model that runs well beats a large one that swaps to disk and takes a minute per reply.
Can I use models I already downloaded with Ollama?
Yes. If Ollama is installed, adris can run models through it rather than pulling a second copy.
How much disk space do models take?
From about 0.7 GB for the smallest usable ones to tens of gigabytes at frontier scale. The catalogue shows the size before you download.
Is a local model as good as GPT or Claude?
No, and anyone saying otherwise is selling something. For summarising, extracting, classifying and drafting they are close enough that you often cannot tell. For hard reasoning, connect your own API key — that option sits right next to the local one.
Does it work without a GPU?
Yes. A GPU makes it faster; it is not a condition of running. The catalogue marks which models are realistic on CPU.
Can I run a model too big for my computer?
Mesh pools memory across machines you own, so two ordinary laptops can run a model neither could hold alone.
Do I need an internet connection?
To download the model once. After that, no.
Related
Run your first local model tonight.
The Models module is free forever, including on the free tier. Pick a model sized for your machine and pull it.