Models

Which model should you pick?

A plain guide to the models you can run in Llamatik: how big they are, how much memory they need, and what each one is good for.

Start here

Any phone, fastest start

Gemma 3 270M (292 MB) answers quickly even on older phones and is multilingual.

Everyday chat on a recent phone

Gemma 3 1B (806 MB) is a good balance. For more depth in many languages, try Qwen 3 1.7B (1.8 GB).

Understand photos and screenshots

SmolVLM 500M (437 MB) describes images and reads the text in them. SmolVLM 256M (175 MB) is the lightest option.

Talk instead of typing

Whisper Base q8_0 (82 MB) is a fast, accurate default. Whisper Small q8_0 (264 MB) copes better with accents and noise.

Chat

ModelDownloadMemory guideGood for
Gemma 3 270M Instruct (Q8_0)292 MB~750 MBFastestMultilingual
Gemma 3 1B Instruct (Q4_K_M)806 MB~1.6 GBBalanced
Qwen 3 1.7B (Q8_0)1.8 GB~3.0 GBBest hereMultilingual
Qwen 2.5 1.5B Instruct (Q2_K)753 MB~1.4 GBMultilingualHeavily compressed
Llama 3.2 1B Instruct (Q2_K)581 MB~1.2 GBMultilingualHeavily compressed
Phi-1.5 (Q2_K)613 MB~1.2 GBHeavily compressed

Vision

ModelDownloadMemory guideGood for
SmolVLM 256M Instruct175 MB~600 MBFastest
SmolVLM 500M Instruct437 MB~1.0 GBBalanced

Voice

ModelDownloadMemory guideGood for
Whisper Tiny (q8_0)44 MB~200 MBFastestMultilingual
Whisper Tiny78 MB~250 MBMultilingual
Whisper Base (q8_0)82 MB~250 MBBalancedMultilingual
Whisper Base148 MB~350 MBMultilingual
Whisper Small (q8_0)264 MB~500 MBBest hereMultilingual
Whisper Medium (q8_0)823 MB~1.4 GBBest hereMultilingualHigh-end devices

Documents

ModelDownloadMemory guideGood for
Nomic Embed Text v1.5 (Q4_0)77 MB~250 MBDocument Q&A

Images

ModelDownloadMemory guideGood for
DreamShaper 8 LCM (Q4_0)1.6 GB~3.4 GBFastest
Stable Diffusion v1.5 (Q4_0)1.8 GB~3.6 GBBalanced
SD Turbo v2.1 (Q4_0)2.2 GB~4.3 GBFastest
DreamShaper XL v2 Turbo (Q4_K_S)2.8 GB~5.2 GBBest hereHigh-end devices

Video

ModelDownloadMemory guideGood for
Wan 2.1 T2V 1.3B (Q4_0)866 MB~2.8 GBNeeds extra files
Wan 2.2 TI2V 5B (Q4_0)3.0 GB~6.0 GBHigh-end devicesNeeds extra files
Wan 2.1 I2V 14B 480P (Q4_0)10.2 GB~16.8 GBHigh-end devicesNeeds extra files

Download sizes are the ones in the app's catalog, which can change over time. The memory figure is a rough guide to the free RAM a model needs while running; real use varies with your device and settings.

What to expect

Models that fit on a phone are small. They answer fast and privately, but they make mistakes, know less about the world than cloud models and cannot search the web. Use them for drafting, summarizing, translating and brainstorming, and double-check important facts. Bigger models answer better but need more memory and storage.

Bring your own model

The app can also import any GGUF chat model from your files or download one from a link, for example from Hugging Face. Larger models give better answers on laptops and desktops with plenty of RAM.

Questions about models

Why is the first download large?
The app does not bundle a model, so it stays small and you choose what to install. A model is downloaded once and stored on your device.
Do I need internet after downloading a model?
No. After the download, chat, vision and voice run offline. You only need a connection to download models and updates.
Which model should I start with?
On a phone, start with Gemma 3 270M or Gemma 3 1B. On a laptop, import a larger model from Hugging Face once you know what you want to use it for.