Choosing a Model

Choosing a Model

Draggy runs two models for conversation, and they do different jobs. Most of the time you only think about the first one. (A third, an embedding model, handles the document library, see Document Library.)

The chat model

This is the one that answers you in the main window and calls tools. It is the big one, and it is what the first-launch download fetches.

Every model on the ladder does the same three things: it calls tools, reads images and reasons before answering. That is deliberate. A model that cannot call a tool closes off half of Draggy, so one is never offered, however small the card.

There are two ladders, because the two kinds of machine do not run the same builds.

Windows and Linux

VRAM	Model
under 4 GB	Qwen 3.5 0.8B
4 GB	Qwen 3.5 2B
5 GB	Qwen 3.5 4B
8 GB	Qwen 3.5 9B
10 GB	Gemma 4 12B
20 GB	Qwen 3.5 27B MoE
24 GB	Gemma 4 31B
32 GB	Qwen 3.5 35B MoE
96 GB and up	Qwen 3.5 122B MoE

Both systems run the same GGUF builds on the same cards, so they get the same ladder. Ollama also publishes nvfp4 and mxfp8 builds, which are faster again, but those need the tensor cores in an RTX 50 series card or newer. Draggy cannot tell which card you have, so picking one by default would strand everybody with an older one. Pull one by hand in Settings → Models if you know your card can take it.

Apple Silicon

Unified memory	Model
8 GB	Qwen 3.5 4B (MLX)
16 GB	Qwen 3.5 9B (MLX)
32 GB	Qwen 3.5 27B MoE (MLX)
36 GB and up	Qwen 3.5 35B MoE (MLX)

On an Apple Silicon Mac, Ollama runs models through MLX, Apple's own framework, rather than the Metal path it used before. It is faster and uses less memory, so Draggy picks the -mlx build of the model rather than the ordinary one.

Two things have to be true before it will. The Mac has to have Apple Silicon, because MLX is Apple's framework for Apple's chips and an Intel Mac has none of it. And Ollama has to be 0.19 or newer, which is the version its MLX engine arrived in. Fail either and Draggy quietly uses the Windows and Linux ladder instead, which runs fine on a Mac, just slower.

The rungs are unified memory, and less of it reaches the model than the number suggests: macOS keeps roughly a third for itself, which Draggy has already subtracted before it chooses. A 16 GB Mac is choosing as though it had about 10 GB, which is why it lands on the 9B rather than something larger.

These are starting points, not rules. Change it whenever you like in Settings → Models, or from the model name at the top of the chat window. The settings panel shows how much of each model will fit in VRAM and how much would spill onto the processor, which is the number that decides whether a model feels usable.

A model that reports "100% GPU" will answer at reading speed. One at 60% will still work, several times slower. Below that it is usually better to step down a size.

Anything with -cloud in its name is hidden. Draggy runs locally by definition. So are models that cannot hold a conversation at all. Draggy asks Ollama what each one can do rather than guessing from its name, and an embedding model is listed but not selectable. Everything else is offered, the smallest models included: on a card that cannot hold anything bigger they are the whole app, and on a larger one they are simply the fast choice.

The talk model

Voice mode uses its own model, and by default not the same one. A spoken reply is two sentences long and has to start arriving within a few hundred milliseconds of you falling silent, which is a completely different problem from writing a good paragraph. The chat model is loaded for context length and reasoning depth, and you pay for both on every turn.

So voice mode picks a smaller one:

VRAM	Model
under 2 GB	SmolLM2 360M
2 GB	Llama 3.2 1B
4 GB	Llama 3.2 3B
8 GB	Gemma 3 4B
16 GB and up	Llama 3.1 8B

The ladder deliberately stops at 8B. Past about four billion parameters a spoken answer is not noticeably better, and the extra weights are paid for on every single turn.

The picker, in Settings → Personalization, has Automatic at the top, which is this table, and below it every model you have installed. If you pick one explicitly it is remembered. If you later delete that model, Draggy quietly goes back to automatic rather than downloading it again behind your back.

When automatic picks something you do not have, it downloads it the first time you press Start talking, with a progress bar and a cancel button. The size is shown under the button before you commit to it.

Running several at once

Ollama keeps models in memory for a while after use. Chat and voice are two different models, and on a card with enough room both sit resident together, which is why replies stay fast when you switch between them.

On a smaller card Ollama will evict one to make room for another, and you will feel that as a pause on the first reply after switching. Picking a smaller talk model is usually the fix.