HT Sherpa

A model is the software that produces AI answers. This guide compares models and tools. New to this? Choose a setup path first. Use the lists below when you want more detail.

New words, in plain language
Local / cloud
Local runs on your computer. Cloud runs on computers managed online.
Model / tokens
A model produces AI answers. Tokens are small pieces of text; some services charge for the tokens used.
Account / identity
An account is a sign-in. Your cloud identity is the account you use for online services.
Connector / permissions
A connector links an AI tool to another service. Permissions are the access you allow.

Field Guide · Week of 5 October 2026

Gemma 4 26B A4B

Google · United States · Google

A larger brain that only wakes a few billion parameters at a time. The best local model on a 32 GB Mac this week.

Where it runs
Runs on your computer
Memory
24–32 GB (4-bit)
Size
26B MoE · 3.8B active
License
Apache 2.0
Hugging Face downloads
13M
Hugging Face likes
1.6k
Card last updated
20 July 2026
On your computer · #1United States · #1

Why it is on this week’s list

Mixture-of-experts means quality without the full 26B tax. Hugging Face downloads put it at the top of the local pile.

Use it when

A 24 GB+ machine and work that needs more range than 12B.

Skip it when

16 GB or less — take the 12B instead.

Published scores

Scores as of 17 August 2026. A blank test is not a zero. We just will not invent it.

GPQA Diamond
82.3*

Google Gemma 4 card · 2026-07-30

SWE-Bench Verified
—

No public number we will stand behind.

LiveCodeBench
—

No public number we will stand behind.

MMLU / MMLU-Pro
82.6*

Google Gemma 4 card · 2026-07-30

MMLU-Pro.

AIME
—

No public number we will stand behind.

Humanity’s Last Exam
—

No public number we will stand behind.

What these numbers actually mean
Walk me through installing thisHugging Face cardAll lists