Choosing a model

Which model fits your graphics card.

Embedding models are listed as well. They turn text into vectors for the Library, and they cannot hold a conversation.

Browsing and downloading#

Settings → Models lists the top 100 trending GGUF models on Hugging Face, with search. Click a model to expand its quantization variants. Each shows its file size.

Click a variant's download button. Two downloads run at a time and the rest wait in a queue, so you can line up as many as you like. The download icon at the right of the search bar shows how many are on their way in a red badge. Click it to see each one with its progress, its time remaining and a cancel button. Cancelling a running download removes the partial .download file immediately, and cancelling a queued one just takes it out of the line.

Each installed model carries tags: Chat, Code, Embedding or Images. Embedding and image models are never offered as a chat model. A model the app cannot run yet, such as speech recognition or text to speech, is hidden from the download list.

Loading into memory#

Draggy loads a model only when you send a prompt. While the engine starts, the status line displays the load progress: copying the model into RAM or video memory, reading the model from disk, reading the prompt, and readying the engine.

Draggy's Models settings with three downloads in progress: two running with time remaining and progress bars, one waiting in the queue, and a red badge showing 3 on the download icon.

Windows and Linux#

Draggy picks the best rung that fits. A dense model must fit in video memory, and never more than 80% of system memory, because the engine reads the whole model into RAM first. A machine with under 2 GB of video memory, such as one with integrated graphics, is treated as having half its system memory instead, up to 8 GB.

Video memory	Dense model
From 1 GB	Qwen 3.5 0.8B
From 2 GB	Qwen 3.5 2B
From 3.5 GB	Qwen 3.5 4B
From 6 GB	Gemma 4 E4B
From 7.5 GB	Qwen 3.5 9B
From 9.5 GB	Gemma 4 12B
From 11 GB	Ministral 3 14B
From 15.5 GB	Qwen 3.8 27B, Q3
From 20 GB	Qwen 3.8 27B
From 23 GB	Qwen 3.8 27B, Q5
From 32 GB	Qwen 3.8 27B, Q8

A mixture of experts (MoE) model keeps its active part on the card and the rest in RAM, so it counts video memory and RAM together: 90% of the first plus 80% of the second must hold the file.

Memory needed	MoE model
14 GB	GPT-OSS 20B
19 GB	Gemma 4 26B
25 GB	Qwen 3.6 35B
40 GB	Qwen 3.6 35B, Q8
61 GB	Qwen 3.5 122B, Q3
82 GB	Qwen 3.5 122B
94 GB	Qwen 3.8 Flash-Next, Q3
116 GB	Qwen 3.8 Flash-Next
135 GB	DeepSeek V4 Flash, Q3
162 GB	DeepSeek V4 Flash
210 GB	GLM-5.3-Flash
250 GB	GLM-5.3-Flash, Q5

The larger ones also ask for a card of 8 GB or more (10 GB from DeepSeek V4 Flash, 12 GB for GLM-5.3-Flash).

Apple silicon#

A Mac has no separate video memory, so Draggy uses 75% of its unified memory and leaves the rest to the system. A base or Pro chip takes the model in the first column. A Max or an Ultra, with far more memory bandwidth, takes the next size up when all of its memory can hold it.

Unified memory	Base or Pro	Max or Ultra
From 8 GB	Gemma 4 E4B	Qwen 3.5 9B
From 16 GB	Ministral 3 14B	GPT-OSS 20B
From 24 GB	Qwen 3.8 27B, Q3	Qwen 3.8 27B, Q3
From 32 GB	Qwen 3.8 27B, Q5	Qwen 3.8 27B, Q5
From 48 GB	Qwen 3.8 27B, Q8	Qwen 3.8 27B, Q8
From 96 GB	Qwen 3.8 27B, Q8	Qwen 3.8 Flash-Next, Q3
From 128 GB	Qwen 3.8 Flash-Next, Q3	Qwen 3.8 Flash-Next
From 192 GB	DeepSeek V4 Flash, Q3	DeepSeek V4 Flash

A Mac with an Intel processor is not Apple silicon. Draggy sizes its models with the rules in the Windows and Linux section above.

Reading the GPU figure#

Settings → Models shows how much of each model fits on the GPU, as a percentage. Green means 98% or more of the model fits. Amber, from 60% to 98%, runs with part of it on the CPU, which is slower. Red, below 60%, is a size too big: 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 model, set in Settings → Chat → Talk.