Weeve now supports Gemma 4, Google's latest open model family for local AI.

That matters because the most important model in Weeve is not running in our cloud.

It is running on your laptop.

That changes the whole product decision. We cannot simply pick the biggest model available and call it done. The model has to be good enough to summarise a real meeting, small enough to fit on a normal Mac, and stable enough to trust every day.

In April 2026, we rebuilt that part of Weeve.

Then we changed the model too.

Local AI has a harder constraint

Cloud meeting tools can hide a lot of complexity.

They can send your transcript to a large remote model, wait for the result, and return a summary. That can work well, but it means your meeting content leaves your device.

Weeve is built differently.

The recording, transcript, speaker detection and summary all run locally. Your meeting content should stay on the laptop where the meeting happened.

That makes the model choice more important.

A local model has to live inside the memory and thermal limits of the machine in front of you. Most newer MacBook Air and MacBook Pro configurations ship with 16 GB of memory. That is the real target, not a benchmark machine in a lab.

The first local model was useful, but not stable enough

Our first small open-weight model did the job.

It fit comfortably on most Macs. It produced useful summaries. It made Weeve feel like the product we wanted: private meeting notes without a bot in the call and without a cloud AI provider in the middle.

But it also crashed.

Not constantly. Not predictably. And not only on smaller machines.

We saw crashes on powerful Macs with plenty of memory too. There was no single root cause we could isolate, fix and move on from. The issue sat in the background for months.

That was not good enough for a tool people use after important calls.

We rebuilt the model runtime on Apple MLX

In April, we stopped treating the crashes as a bug to patch around.

We rebuilt the way Weeve runs local models on the Mac.

The new foundation is based on

Apple MLX, an open-source framework designed for machine learning on Apple silicon and optimized for Apple's unified memory architecture.

That mattered because the old plumbing was where the crashes lived. Once we replaced it, the failure path disappeared with it.

The rebuild also let us remove the optional cloud summarisation path.

That cloud option had been off by default. Users could connect their own Anthropic, Google or OpenAI account if they wanted to. Almost nobody did.

That told us something simple: people were choosing Weeve because they wanted meeting summaries on-device.

So we made the architecture match the promise.

Why Gemma 4 was worth switching for

After the rebuild, we briefly ran a Qwen model on the new runtime.

It worked.

Then Google released Gemma 4.

Google describes Gemma 4 as its most capable open model family to date. The family includes E2B, E4B, 26B Mixture of Experts and 31B Dense variants, which gives developers a cleaner path from edge devices to larger local workstations.

The version that mattered most for Weeve was Gemma 4 E4B.

The official model card lists Gemma 4 E4B at 4.5B effective parameters, with a 128K context window and native support for text, image and audio. That combination matters for meeting notes because long context and efficient local inference are both more useful than raw size alone.

The benchmark picture also made the switch easier to justify. In Google's published results, Gemma 4 E4B beats Gemma 3 27B on several reasoning and coding-oriented benchmarks despite being much smaller. It scores 69.4% on MMLU Pro versus 67.6% for Gemma 3 27B, 42.5% on AIME 2026 versus 20.8%, and 52.0% on LiveCodeBench v6 versus 29.1%.

Those benchmarks are not the same as summarising a meeting. But they do show the pattern we care about: better capability per local compute budget.

An independent arXiv evaluation reached a similar practical conclusion. It found that Gemma 4 E4B stayed close to the top models across prompting settings while using substantially lower latency and memory, making it a strong real-world operating point rather than just a leaderboard entry.

That is the kind of model Weeve needs.

Not the biggest model. The best model that can run privately on the machine people actually use.

What runs on your Mac today

Weeve now picks the model that fits your machine.

Smaller Macs get a smaller model. Most users with a 16 GB Mac land on Gemma 4 E4B, which gives us the best balance of quality and local performance today.

If your Mac has more memory, Weeve can offer a larger model. That model is a little slower, but it can be better on long or dense meetings.

You can override the recommendation in settings.

The one rule Weeve enforces is simple: it will not try to load a model your Mac cannot actually run.

Why this matters

This was not a benchmark chase.

For Weeve, a better local model means better meeting summaries without changing the privacy model.

Your audio does not need to go to us. Your transcript does not need to go to us. Your summary does not need to go to a cloud AI provider.

The product gets better while the meeting stays local.

That is the point.

What comes next

The next step is not just a stronger general model.

Once we have enough opted-in feedback, we want the local model to get better at the kind of summaries Weeve users actually want.

Not a generic meeting recap.

Your style of summary. Your level of detail. Your way of turning a messy conversation into follow-up, decisions and next steps.

That is where on-device meeting intelligence becomes more interesting than the cloud version.

Not just as good.

More personal.

And still private by design.

Sources

Google: Gemma 4, byte for byte, the most capable open models

Google AI for Developers: Gemma 4 model card

arXiv: Unified Deployment-Aware Evaluation of Open Reasoning Language Models