Taby's local brain
You don’t choose models, sizes or settings. Taby picks, sets up and tests the model, then runs it on your computer. Taby’s Local Brain is Qwen 3.5 4B, an open model from Alibaba. Taby picked it because it scored highest on Taby’s own test of 114 real tasks: about 96 of 114, against about 72 for Gemma 4 E2B. It’s mostly better at actually doing things, like moving, snoozing and completing tasks. Taby tuned it for everyday jobs like tasks, reminders and notes.
Updated October 3, 2026
The Local Brain
Taby tested small open models on 114 real tasks, like adding, moving, snoozing and completing things. Qwen 3.5 4B came out on top.
| Model | Score of 114 | In plain words |
|---|---|---|
| Qwen 3.5 4B | 94–101 | Best at actually doing things: moving, snoozing, completing and deleting items. No broken actions in any run |
| Gemma 4 E4B (today’s Standard) | 77–98 | Close on a good run, less steady, 1 to 5 broken actions per run |
| Gemma 4 E2B (today’s Light) | 68–81 | Misses more of the doing, 1 to 5 broken actions per run |
Ranges over several runs of the same test.
Before the Local Brain
Earlier versions offered four sizes. The update moved everyone to Qwen 3.5 4B in the background.
Light
Gemma 4 E2B
3.35 GB model
Standard
Gemma 4 E4B
about 6.1 GB model
Large
Qwen 3.5 9B
5.9 GB model
Extra large
Gemma 4 12B
7 GB model
The B means billions of learned parameters, approximately. A larger parameter count gives a model more capacity, but it also asks more from your computer
Model size is not the whole answer. How a model was trained, how it is compressed, how long the chat is, and how much memory is free all change how fast and useful it feels. That’s why Taby tests models on real tasks instead of going by size
Taby's models are built to run on your computer, so everyday assistant work can stay local. Optional cloud AI is available for harder jobs when you choose to use it
These are language models. Voice input and spoken replies also depend on Taby's separate speech-to-text and text-to-speech components
New to this? Local AI, explained in plain words, or see how Taby compares with Ollama and OpenClaw.
Hardware fit
Taby’s local AI needs 8 GB of memory or more (a graphics card on Windows). The free browser test checks your computer in a few seconds. If it can’t run the local AI, Taby Cloud does the AI part instead
Test this computerQuick answers
One Local Brain for everyone: Qwen 3.5 4B, an open model from Alibaba. Taby downloads it, sets it up and keeps it up to date.
Because it did best on Taby’s own test of 114 real tasks: about 96 of 114, against about 72 for Gemma 4 E2B. It’s mostly better at actually doing things, like moving or completing a task, instead of only answering.
Yes. Taby tuned it for everyday jobs like tasks, reminders and notes.
No, and that’s on purpose: Taby picks, sets up and tests one model so you don’t have to. If you want to try lots of models, apps like Ollama or LM Studio are made for that.
Use Taby Cloud. It runs the AI for you, so it works on smaller laptops and on Windows PCs without a graphics card.
No. The model writes Taby’s answers. Listening and talking use separate speech parts that also run on your computer.
What's next
Use one ready-to-go local assistant, or choose a model file and build the setup yourself
Ready to use
Taby is the all-in-one local option: the model, chat, tasks, notes, planning, and voice work together in one assistant on your computer
See what Taby can doDIY setup
Choose a GGUF model file on Hugging Face, then run it locally with Ollama, LM Studio, or llama.cpp. You manage the runtime and updates
Browse GGUF models