Pretype

Pretype is an open-source autocomplete for macOS. It suggests the next words in any app and fixes typos as you type. The models on this page run on your Mac with MLX.

We measure every model on the same held-out eval: 5,649 rows of real human text in 17 languages, plus separate sets for Arabic, Hebrew and Danish. See how to choose a model.

Our own model

Model	Size	What it does
EuroLLM-1.7B-Pretype-5bit	1.1 GB	One file completes text and fixes typos. It fixes more than Gemma 4 E2B 4-bit and touches less clean text, with 33% less energy. It writes no replies.

Completion models

Pretype selects the default from the memory of your Mac.

Model	Size	Use
gemma-4-e4b-6bit-slim	5.3 GB	Default from 32 GB
gemma-4-e4b-8bit-slim	6.3 GB	Manual pick for the best quality
gemma-4-e2b-8bit-slim	3.6 GB	Default at 24 GB
gemma-4-e2b-4bit-slim	2.6 GB	Default at 16–18 GB
Qwen3.5-2B-M40-6bit	1.5 GB	Default at 8 GB. One file completes text, fixes typos and writes replies.
Qwen3.5-2B-Base-6bit	1.5 GB	The best small Base build for many languages
Qwen3.5-2B-Base-4bit	1.1 GB	The same model at 4 bits
MiniCPM5-1B-Base-8bit	1.1 GB	Fast, for English

Fix and reply models

Pretype loads these models on the first fix or reply.

Model	Size	Used by
gemma-4-e4b-it-6bit-slim	5.5 GB	Gemma 4 E4B
gemma-4-e4b-it-4bit-slim	4.5 GB	Gemma 4 E2B 8-bit
gemma-4-e2b-it-4bit-slim	2.9 GB	Gemma 4 E2B 4-bit
Qwen3.5-2B-4bit-slim	1.1 GB	Qwen3.5 2B Base

About these builds

The -slim models are re-packs without the audio and vision weights. All other weights do not change, and the text output matches the original.
The other models are our own MLX conversions, merges or fine-tunes. Each model card gives the source, the recipe and the measurements.
Each model keeps the license of its source. See the model card.

models 13

datasets 0

None public yet

AI & ML interests

Recent Activity

Team members 1