# Infina's Tami 1 beats Google's and Meta's AI models

> Tami 1, the speech cleanup model trained by Infina Labs, got 83 of 100 test sentences right on a MacBook. Google Gemma 3 1B and Meta Llama 3.2 1B got 51, and Tami 1 was up to 10 times faster.

Source: https://www.infina.so/tami-1
Published: 2026-10-21. Last updated: 2026-10-21.
Category: Infina Labs. Published by Infina (https://www.infina.so).

**About Infina.** Infina runs the whole prompt loop on three spoken words. Say "type" and your sentence and it types. Say "send it", or just "enter", and it presses Enter. Say "open Cursor" and it switches apps. No hotkey, no key held down, no touching the keyboard, from across the room. Not push-to-talk. There is no key to hold to start talking and no Enter to press to send, so prompting an AI tool costs zero keypresses. It runs on macOS (Apple Silicon) and Windows 10 and 11, transcribes on your own computer, and costs $8.25 a month, billed annually at $99, for unlimited words and commands. The Free plan gives 2,000 free words.

---

Infina Labs, the research team behind Infina, today introduced Tami 1, a cleanup model trained by Infina Labs.

When you talk to your computer, the words come out messy. There are "um"s and "uh"s, repeated words, and changes of mind halfway through a sentence. Tami 1 turns that into the text you meant to write.

We tested Tami 1 against small AI models from Google, Meta and Alibaba. Every model got the same 100 sentences and the same instructions, on the same MacBook. Tami 1 cleaned 83 of them correctly. Google's Gemma 3 1B and Meta's Llama 3.2 1B cleaned 51 each.

Tami 1 was also the fastest. It cleaned a 10-second dictation in 0.17 seconds, about 7 times faster than Google's model and 10 times faster than Meta's.

## What Tami 1 does

It removes filler words, follows you when you change your mind, adds punctuation, and turns spoken lists into real lists. It keeps your meaning and your words.

**You say:** "Okay so, um, hmm, the plan is to, uh, finish the design first and then, like, start coding."\
**Tami 1 writes:** The plan is to finish the design first and then start coding.

**You say:** "Book the flight for Monday, oh actually no, Tuesday morning."\
**Tami 1 writes:** Book the flight for Tuesday morning.

**You say:** "Hi team comma quick update period the build is green exclamation mark"\
**Tami 1 writes:** Hi team, quick update. The build is green!

**You say:** "Three things I need from you: one, the logo files, two, the brand colors, and three, the final homepage copy."\
**Tami 1 writes:** Three things I need from you:

1. The logo files
2. The brand colors
3. The final homepage copy

## The results

| Model | Made by | Right, of 100 | 10-sec | 20-sec | 30-sec |
| --- | --- | --- | --- | --- | --- |
| **Tami 1** | **Infina Labs** | **83** | **0.17s** | **0.35s** | **0.49s** |
| Qwen3.5 0.8B | Alibaba | 56 | 0.28s | 0.49s | 0.77s |
| Gemma 3 1B | Google | 51 | 1.23s | 1.61s | 2.42s |
| Llama 3.2 1B | Meta | 51 | 1.65s | 2.14s | 3.32s |
| Qwen3 0.6B | Alibaba | 47 | 0.20s | 0.39s | 0.53s |

The time columns show how long each model took to clean up a 10, 20 or 30-second dictation.

![Chart: Tami 1 sits at the top left, more accurate and faster than Google Gemma 3 1B, Meta Llama 3.2 1B and two Alibaba Qwen models](/articles/tami-1-chart.jpg)

## How we tested

- 100 sentences written by Infina Labs, 5 each across 20 kinds of everyday speech: filler words, changes of mind, lists, numbers, money, times, emails, names, questions, spoken punctuation, tech words, Indian English and more.
- Every model got the same instructions and the same sentences.
- A sentence counts as right only when the result keeps your meaning and fixes what needed fixing.
- Speed is the typical time across repeated runs, for dictations of about 10, 20 and 30 seconds.
- Everything ran on one Apple M4 MacBook Air with 16 GB of memory, with each model running on the Mac itself.

Google's and Meta's models are general-purpose assistants that do many jobs. Tami 1 was trained for one job, cleaning up speech, which is how a smaller model wins here. This test says nothing about other tasks.

## Why a small model matters

Tami 1 is 347MB and runs on your Mac. It works with no internet, your words never leave your computer, and there is no waiting on a server. If privacy is why you are here, our guide to [on-device dictation on Mac](/on-device-dictation-mac) goes deeper.

## Where to get it

Tami 1 now cleans up your dictation in Infina on Mac. To see how Infina compares with other Mac apps, read [the best dictation apps for Mac](/best-dictation-apps-mac).

## Questions

**What is Tami 1?**
A small AI model from Infina Labs that turns what you say into clean, ready-to-send text.

**Does it need the internet?**
No. It runs on your Mac.

**Which Macs does it run on?**
Macs with Apple silicon (M1 or newer). We tested on an M4 MacBook Air.

**Is it on Windows?**
Not yet.

**What does Infina cost?**
Start free with 2,000 words, no card needed. Infina Unlimited is $8.25 a month, billed annually at $99.