01What Is Gemma 4 90% Faster and How It Works

Gemma 4 90% faster is a free AI model from Google that now runs much faster on your Mac. It uses MLX and Ollama to achieve up to 90% speed improvement over previous local versions.

Before this update, local models were slow and not very useful. Now with MLX, Gemma 4 can run at 95 tokens per second instead of 50. That is a big jump.

MLX works by auto-tuning how many tokens to guess ahead. It drafts three to five steps and keeps the right ones. This speeds up the output without slowing down.

You can run Gemma 4 completely free on your own machine. No cloud costs, no internet needed after download. The model is designed for mobile devices, so it works well on a Mac Studio or even a laptop.

02Main Features of Gemma 4 90% Faster

  • Free and local: No subscription or cloud service required.
  • Up to 90% faster: Speed increase from 50 to 95 tokens per second in tests.
  • MLX integration in Ollama: You can download MLX models directly through Ollama without extra setup.
  • Auto-token guessing: The model guesses ahead to generate text faster.
  • Works on Mac: Optimized for Apple Silicon, but runs on any Mac with Ollama.
  • Open source: The MLX project is on GitHub and free to use.

This update makes Gemma 4 the first local model that is actually useful for everyday tasks. It is fast enough to build apps, write content, and automate simple jobs without waiting long.

03How to Set Up Gemma 4 90% Faster on Your Mac

  1. Install Ollama. Go to ollama.com and download the free app for Mac. Install it like any other program.
  2. Update Ollama. Open terminal and run ollama pull gemma:2b-mlx or the command for the MLX version. Make sure you have the latest Ollama version.
  3. Run the model. Type ollama run gemma:2b-mlx to start. It will run locally on your Mac.
  4. Optional: Use with agent OS. If you have the agent operating system, plug Gemma 4 into it as your local engine. That lets you build apps and get previews automatically.

That is all you need. The whole process takes a few minutes. You do not need a powerful GPU.

A Mac with Apple Silicon works best.

04How to Use Gemma 4 90% Faster: Examples Walkthrough

In this episode, Julian Goldie showed several examples of using Gemma 4 with the agent operating system. First, he built a to-do list app. He typed 'create a to-do list app' and the model coded it in the background.

Then a preview appeared. The app worked: he could add and delete tasks.

He also built a game called Dragon Realm. The output looked impressive, with a nice interface. For a landing page, the results were less visual but still functional.

The model handled basic builds well.

He used Hermes agent to run Gemma 4 inside the agent OS. The system saved all builds to a workspace. You can iterate as many times as you want.

The model runs 24/7 if you leave it on.

For coding tasks, Gemma 4 was about 60% faster than the old version. It handled prompts quickly and output code without long waits.

05Results and Speed Tests of Gemma 4 90% Faster

Julian tested the speed using Claude to measure it. The old version ran at 50 tokens per second. The new MLX version ran at 95 tokens per second.

That is a 90% increase in peak speed.

In real use, most tasks were about 60% faster. That still makes a big difference. You can see the model finish prompts much quicker than before.

He also compared the same prompts on the old build (using llama.cpp) versus the new MLX build. The MLX build was clearly faster. The improvements come from MLX's auto-token guessing, which speeds up generation without losing quality.

The builds he created worked correctly. The to-do list app was functional. Dragon Realm looked good for a local model.

So the speed gain does not hurt output quality for basic jobs.

06Tips, Limits, and Mistakes to Avoid with Gemma 4 90% Faster

  • Do not use it for graphics-heavy projects. The model struggled with visual landing pages. It works best for text and simple UIs.
  • It is not frontier-level. Gemma 4 is free and local, so it won't match top models like Fable 5 for hard tasks. Use it for small builds and content.
  • Make sure you have the latest Ollama. Old versions may not support MLX models. Update via terminal before pulling.
  • Iterate often. You can run the model 24/7. Keep asking it to improve the output until you are happy.
  • Use for everyday automation. Writing blog posts, basic coding, and simple scripts are perfect for this speed.
  • Do not expect 90% every time. The 90% figure is peak. In tests, it was about 60% faster on average. Still good.

If you follow these tips, Gemma 4 becomes a reliable free tool for your daily workflow.

07Who Gemma 4 90% Faster Is For and Why It Matters

Gemma 4 90% faster is for anyone who wants a free local AI that actually works. Coders can use it to build apps without paying for cloud APIs. Non-technical users can run it with the agent operating system and still get results.

It matters because local models were too slow to be useful until now. This update changes that. You can run it on a Mac, it is free, and it is fast enough for real work.

Julian Goldie says it is the first time local models have done something decent and useful.

Even if you are not technical, you can learn to use it. The agent OS makes it simple. You can ask Gemma 4 to build a to-do list and see a preview.

That opens up AI automation to more people.

If you have a Mac and want a free local AI for everyday tasks, Gemma 4 90% faster is a good choice.