Draggy
Getting started
Choosing a model
Which model fits your graphics card, and the one number that decides it.
Draggy asks Ollama what each installed model can do rather than guessing from its name. Cloud models are hidden; embedding models are listed but cannot answer.
Windows and Linux
Picked from the graphics card's memory first. A machine with little or none still gets a model sized to what its system memory can hold, rather than the smallest one there is.
| VRAM | Model |
|---|---|
| under 4 GB | Qwen 3.5 0.8B |
| 4 GB | Qwen 3.5 2B |
| 5 GB | Gemma 4 E4B |
| 6 GB | Mistral 7B |
| 8 GB | Qwen 3.5 9B |
| 10 GB | Gemma 4 12B |
| 12 GB | Phi-4 14B |
| 16 GB | GPT-OSS 20B MoE |
| 20 GB | Qwen 3.5 27B MoE |
| 22 GB | GLM 4.7 Flash MoE |
| 24 GB | Gemma 4 31B |
| 32 GB | Qwen 3.5 35B MoE |
| 96 GB | Qwen 3.5 122B MoE |
Every model this can name, with its parameters, best use and what it supports, is in the full model table.
Apple Silicon
Ollama 0.19+ runs models through MLX, so Draggy picks the -mlx
build. There is no separate video memory on a Mac, so the sizes below are the
whole, unreduced number a Mac shipped with. A base or a Pro chip takes the
model in the left column; a Max or an Ultra, with far more memory bandwidth,
takes the one beside it.
| Unified memory | Base or Pro | Max or Ultra |
|---|---|---|
| 8 GB | Qwen 3.5 2B (MLX) | Qwen 3.5 4B (MLX) |
| 16 GB | Qwen 3.5 4B (MLX) | Qwen 3.5 9B (MLX) |
| 24 GB | Qwen 3.5 9B (MLX) | Qwen 3.5 27B MoE (MLX) |
| 32 GB | Qwen 3.5 27B MoE (MLX) | Gemma 4 31B (MLX) |
| 48 GB | Gemma 4 31B (MLX) | Qwen 3.5 35B MoE (MLX) |
| 64 GB | Qwen 3.5 35B MoE (MLX) | GLM 4.7 Flash (MLX) |
| 96 GB | GLM 4.7 Flash (MLX) | Qwen 3.5 122B MoE (MLX) |
| 128 GB and up | Qwen 3.5 122B MoE (MLX) | Qwen 3.5 122B MoE (MLX) |
Every Mac able to run its own models, since the first one in 2020:
| Mac | Years | Chip | Unified memory |
|---|---|---|---|
| MacBook Air | 2020–2025 | M1, M2, M3, M4 | 8–32 GB |
| MacBook Pro 13" | 2020 | M1 | 8–16 GB |
| MacBook Pro 14"/16" | 2021–2025 | M1 to M4, Pro or Max | 16–128 GB |
| Mac mini | 2020–2024 | M1, M2 or M2 Pro, M4 or M4 Pro | 8–64 GB |
| iMac | 2021–2025 | M1, M3, M4 | 8–24 GB |
| Mac Studio | 2022–2025 | M1/M2 Max or Ultra, M3 Ultra, M4 Max | 32–512 GB |
| Mac Pro | 2023 | M2 Ultra | 64–192 GB |
A Mac from 2019 or earlier is Intel: it has a real, separate graphics card instead, and reads the Windows and Linux table above.
The same table, sized for unified memory instead of VRAM, is the full model table for Mac.
The number that matters
Settings → Models shows how much of each model fits in video
memory. At 100% on GPU it answers at reading speed. At 60% it works, several
times slower. Below that, step down a size.
A loaded model gets an Unload button there, which frees the video memory without deleting anything.
The talk model
Voice mode uses its own, smaller model, set in
Settings → Chat → Talk. Answering out loud is worth more than
answering well, so keep it small.