When ChatGPT first launched in late 2022, it was a hit, and months later it nearly broke the internet. It was a model that could handle queries, write for you, have conversations and help you make decisions, none of which existed before on this scale. Later competitors drove the cost even further, and within a few years AI stopped feeling futuristic and became part of the new normal.
So, naturally, there’s something revolutionary about running a model of similar reasoning on a laptop with no internet connection at all, running at thirty thousand feet on an eight-hour flight. 16 GB of RAM. Gemma 4 has become undisputed the most popular domestic model today, in part because it doesn’t differentiate between the devices it runs on. It can be your flagship workstation or budget gaming laptop, and its use cases are only limited by your imagination. Here’s why I can’t stop using it.
16 GB of memory and almost no compromise in AI experience
Gemma 4 gets it right like no other model does, and there’s a clever trick behind it
Now, I’m not a complete stranger to the on-premises AI market, having tested every model over time that doesn’t require data center-level computing. When you test drive multiple models, you’ll realize that compact sizes always come with compromises. This could be factual depth, operational relevance, or relevance in longer conversations. Gemma 4 E4B variant feels like a rare departure from this pattern, so the model somehow doesn’t feel like it comes with the usual set of compromises, and that’s because of an architectural decision on Google’s part.
To understand this, it is necessary to study the nomenclature more deeply. The “E” in the model’s name stands for “effective,” and that means something when it comes to LLMs. The E4B variant actually stores 8 billion total parameters, but only 4.5 billion of them are activated at once. The rest resides in what Google calls “Allocation Per Layer”, where each decoder layer gets its own layout table for its own little token.
These are large in storage but ultimately cheap, they work more like fast searches than continuous computation. This means that the model runs with the memory space and speed of the 4B model and uses nearly twice as much stored “knowledge”. On the other hand, competing models in this size spend all their parameter budget on tight weights and have nothing to spare.
The use cases keep finding me, not the other way around
A view for my notes, storage for my archives, and tools I’m just starting to discover
I strongly feel that anyone who hasn’t used Gemma 4 offline can’t fairly appreciate its raw capabilities and usability. It appears as a small model with modest features, and every instinct of a user accustomed to using cloud AI will tell them that the 4B model will be like a mindless chatbot.
Then, you realize that you can give him a photo of your terrible handwriting and watch it transcribed, structured, and ready to be hidden away in his knowledge base, defying every notion of the ability that exists right now. This is of course just one example of many using the model native vision abilities. You can display it on labels, manuals, recipes, legal documents and watch it mean everything to you.
If that hasn’t sold you on the idea of running the model yourself, then there are cases I wouldn’t use. The Gemma 4 E4B supports native function calling, and perhaps that’s the feature that’s featured here more than anything else. You can describe small tools to the model, ask it to “look in the display” or “read this file,” and instead of responding in text, Gemma decides what work is required and when, fills in the arguments, and hands off the execution to the software on your machine.
One popular use case for this is through Obsidian, which becomes a completely private, offline second brain using Ollama as AI Providers or local runtimes and community plugins like Native GPT to combine the two. If you show the setup at checkout, the model can summarize, rewrite, and answer questions with your notes as context. The model controls the decision-making, the machine controls the execution, and your archive starts talking to you.
No counters, no levels, no usage limits, and no accounts
When it comes to AI, the consumer market is accessible to almost anyone with an internet connection, but the most advanced capabilities are kept behind a premium subscription tier. Furthermore, every cloud AI service, no matter how generous, scales at some point somewhere. Tokens, message caps, priority queues, “advanced” models one step up… the pattern is familiar.
In the early days, the appeal of native AI was that it was free. The appeal for me now is that it’s limitless, accessible from anywhere, and doesn’t have pop-up alerts telling me I’ve exceeded my usage quotas. Once the weights are on the SSD, the query only takes a few seconds and nothing else. It simply changes the way I behave and interact with AI tools. I iterate drafts ten times instead of twice, generate the same code multiple times until I’m satisfied, and query my knowledge base as much as necessary to develop a working understanding.
The best AI tool on my laptop is the one no one gives me credit for
There are many open-weight models out there, but the capabilities of the Gemma 4 E4B make every one of these tasks easy, as it is the sweet spot between efficiency and capability, if ever there was one. It requires nothing more than 16 GB and a bit of disk space, and returns a solid helper without adding a single line (or step). Three months from now, it will cost me the same as it did on day one, which makes it as reliable as it gets.
