Today I’m releasing a beta of Micro.blog 4.0 for Mac. It’s a big update and I’d like to get some feedback before finalizing it. You can download the beta here.

The first thing you’ll notice is that we’ve updated the design with a bunch of little improvements. The sidebar and toolbar better fit with modern macOS. Spacing has been tweaked and UI quirks have been fixed. Blog category editing has been added.

I’ve been using this version all week and it’s much better. It’s not a redesign. More of a fine-tuning.

The other big change is support for local AI running on your Mac. In the preferences window, you’ll find two checkboxes:

• A global opt-out checkbox that stops Micro.blog servers from using AI anywhere for your account. That checkbox has been on the web for a while. Now it’s conveniently on the Mac too.
• A new checkbox to enable local AI models. This is only available on M1 or later Macs with at least 24 GB of RAM. The requirements are high because I didn’t want local AI to be a step back in quality.

When enabling local AI, Micro.blog will download Gemma 4 26B. The model takes up about 14 GB on disk. The feature is off by default and completely optional. If you enable the feature and then later disable it, Micro.blog removes the local files.

To avoid problems with increased memory usage, we only load the AI model when needed. When Micro.blog goes to the background, even if the app is still running, we unload the model. My goal is for you not to notice it’s there.

For the initial release of 4.0, local AI use is fairly minimal. When attaching a photo to a blog post, if you click on the photo, it will use the local model to generate accessibility text that you can edit. Without local AI, Micro.blog will upload the photo to its servers to generate accessibility text.

The feature is the same whether you have local AI enabled or not, so you aren’t missing anything if your Mac doesn’t meet the requirements. The only difference is where the AI runs.

You’ll also notice some text in the preferences window:

Your Mac can be used with local AI models when possible for your own data, as well as to process public posts that are available to multiple Micro.blog users, reducing reliance on cloud AI services for everyone.

What does that mean? This is an experiment to take some of the work that would usually be sent to centralized services like OpenAI and instead offload it to Macs. For example, we spend a lot of time just summarizing blog posts, or figuring out the best keywords for them. This is data that a Mac could process as it sits there idle.

AI is a divisive topic, and a special point of contention is data centers. By leveraging Macs, we could reduce some of the dependence on data centers, effectively using a little bit of energy at the edge of the network, in a way that benefits all users.

Here’s how this might work in practice:

• Let’s say you are a Micro.blog Premium customer who uses Inkwell’s Reading Recap feature.
• For the feeds you subscribe to, there would occasionally be new posts that need to be processed. Some of them would be sent to your Mac, where local AI would run.
• The result would be contributed back to the platform, where the data could be surfaced for other users who subscribe to those same feeds.

There might be other uses, not just for the feeds you subscribe to, but it would always follow this pattern of only using public blog data that is the same for multiple users, to avoid leaking any private information.

We’re going to go slowly. While the checkbox is present in the current build, the server-side work is not yet enabled for any users. The code is also careful to only run when your Mac is plugged into power and idle. The client code is open source so it can be inspected.

Thanks for your support! I’m happy to hear feedback during the beta, and we’ll adjust as necessary.