Offline models

Offline models run entirely on your machine. Download once, use forever, free - and nothing you type ever leaves your computer.

Best for this computer

The models page opens with one fit-checked pick per activity - codes with you, everyday chat, sees images, health questions - each chosen from what your machine can actually run, with download in place. The health pick can become your health-questions model in one tap.

The catalog

Every model in the catalog - offline and online - has its own page at yourownai.net/models, with sizes, context, license, and a check of what runs on your machine using the same sizing the app uses. This page explains what the app shows you.

Offline Models lists a curated catalog with filters - coding, reasoning, agentic, writing, math, vision - and sort options. Each variant shows its download size, parameters, context window, quantization, and memory needs, plus the fit verdict for your machine:

• Full speed - fits fully on your graphics card.
• GPU + RAM - a mixture-of-experts model bigger than your graphics memory: its rarely-used parts stay in main memory, the rest runs on the card. Fast for its size - a 35B model can run at a useful speed on an 8 GB card with 32 GB of main memory. The badge's hover shows the split in use ("the experts of 37 of 40 layers in main memory"). The app checks that the model's experts really are rarely used before making this promise: a converted model that fires most of its width on every answer is graded "Runs slower" instead - it still runs if you pick it, but the badge won't claim speed the hardware can't deliver.
• Runs slower - on a machine without a usable graphics card, a model that is tight in main memory.
• Too large - will not run on this device. A dense model that does not fit the graphics card is too large here: it loads fully on the card or not at all.

A model file the engine cannot read is said in those words - in chat and on its card - never as "too large for your graphics card". A vision model is graded for text on its own: its vision add-on loads only when you send an image, and the card says when the add-on would not fit alongside right now (see Images).

Some mixture-of-experts models come in two builds: the regular one, and an expert-lean build that squeezes the parts kept in main memory so the same model reaches machines with less of it (Qwen 3.6 35B and Gemma 4 26B each have a build for 24 GB machines).

Every card says who made the weights, who packaged the files you download, and - for community builds - what they are based on and how they differ: "Official Google weights, packaged by Unsloth" or "Community build by empero-ai - distilled from Qwen 3.8". You always know whose work you are running.

Licensed models (like MedGemma, Google's medical model) add a short agree-to-the-publisher's-terms step in the download flow.

Beyond the catalog

Add Custom Model from Hugging Face takes any GGUF model URL. The System Information block shows your RAM, GPU, VRAM, and the models directory.

Managing models

Where models are stored. Models and the app's other large downloads (vision files, speed-up files, engine packs) live in one folder. A Change control - on this page and in Settings > Components - moves everything to a folder you pick on any drive, with live progress; downloads follow the new location, and a move that fails part-way rolls back completely. Every download also checks free space up front and says plainly what is missing ("needs 22.4 GB free, 9.1 GB available") instead of failing mid-file.

Downloaded models sit in a compact list - fit badge, the context each runs at on this machine, its measured speed here once you have used it ("~27 tok/s measured"), pause, delete - with your disk total in the header. Pause a model to hide it from pickers without deleting it. Deleting a model that an AI uses tells you which AIs are affected and what happens to them. Downloads resume if interrupted, survive navigating away, and pick themselves back up after the app was closed mid-way. A file that is incomplete or corrupted is listed as damaged, never offered, and downloading the model again replaces it. When a maker replaces a file after the catalog pinned it, the row says Re-download needed instead of a fit badge - the catalog now points at the copy that loads, with delete and download one click away.

Fine-tune. Every downloaded model has a Fine-tune button on its row: context size, expert layers in main memory, the speed-up file - for this computer only, automatic until you set a number, and a measured tune-up if you want the real numbers. A tuned model says "fine-tuned" on its row. See Fine-tune.

Very large models (the workstation class - DeepSeek V4 Flash, GLM-5.2) ship as several files. The app downloads the parts in order, shows "part 2 of 3", and treats the set as one model. They stay off the page on machines that cannot run them.

Speed-up files. Where the maker ships one (Nemotron 3.5 Lightning's multi-token head, DeepSeek V4 Flash's DSpark draft), the app downloads it after the model and keeps it beside it; the engine then checks several drafted tokens per pass and answers faster - most of all when a model's experts live in main memory. A model downloaded earlier offers "Get its speed-up file" on its row.

Each model file is fingerprinted once in the background, and every answer it gives is recorded with that fingerprint - see Your records.

Health stays home

Health questions are answered on your device by default, by the model you choose for them - and online only as an explicit, asked-every-time choice. The whole story: Health questions.