DS4 by antirez, the artificial intelligence that runs at your home

There is a Sicilian programmer, Salvatore Sanfilippo, whom everyone in software knows as antirez. He is the man who invented Redis, one of those invisible pieces of infrastructure that keep half of the internet running. Many also know him for a masterpiece of code called dump1090, which lets you "intercept" the squawk messages of air traffic using a usb device costing just a few tens of euros. In May 2026 he released something small and sharp that, in my opinion, is worth understanding. It is called DS4, and it makes a simple promise. Take one of those big artificial intelligence models, the kind that usually only runs inside datacenters, and make it run on your own computer. At home, in the office, offline, without asking anyone for permission.

Let us try to explain what it actually does, why it is clever from a technical standpoint, and above all why we think it is a piece of news that speaks to something far bigger than software.

First of all, what does it mean to run a model

When you use an AI assistant on a website, you are not using anything on your own computer. Your text leaves, travels to a datacenter on the other side of the world, where a cluster of extremely expensive graphics cards computes the answer, and it gets sent back to you. The piece of software that does that computation is called an inference engine. It is the engine that, given the model and your question, produces the words of the answer one after another.

DS4 is exactly this, an inference engine. The difference is where it runs. Not in a datacenter, but on a machine you can keep on your desk.

The model it runs, DeepSeek V4 Flash

DS4 was built for one specific model, DeepSeek V4 Flash, released as open source in April 2026 by a Chinese lab. On paper it is a huge model, 284 billion parameters, but it is built around an idea called Mixture of Experts. In plain terms, the model is divided into many small specialists, and for each word it generates it lights up only a handful of them, the right ones for that piece of reasoning. The others stay switched off. This property is called sparsity, and it is the reason a gigantic model can still run without a datacenter underneath it.

This is no small detail. It means knowledge is spread across many experts, and in every single step most of those experts are not even touched.

The first trick, compress only where it does not hurt

A model of this size, kept at full precision, takes up an amount of memory that no normal computer has. The classic solution is called quantization, meaning you reduce the precision of the numbers that make up the model. It is a bit like going from a photo with millions of colors to one with far fewer colors, it takes up less space, but if you overdo it, it shows.

Here antirez makes the smart move. Instead of compressing everything the same way, he compresses asymmetrically. The Mixture of Experts experts, which as we said mostly stay switched off, are crushed down to 2 bits per number, a brutal compression. Everything else, the attention, the router that decides which experts to activate, the delicate parts, stays at full quality. The result is that the model drops to about 76 gigabytes, fits into a machine with 128 gigabytes of memory, and the quality holds. It holds partly because the model had been trained on purpose to tolerate this compression, so the credit is shared between the people who built DeepSeek and the people who built DS4. On this point antirez is honest and says so plainly.

The KV cache, the model's working memory

Now we get to the heart of the matter, and it helps to understand something that usually stays hidden.

When you talk to a model, it does not reread the whole conversation from scratch for every word. That would be painfully slow. Instead, while it reads what you wrote, for each piece of text it jots down internal notes, a kind of shortcut it uses so it does not have to recompute everything every time. These notes are called the KV cache, the cache of keys and values. You can think of it as the model's working memory, the scraps of paper on which it has written down everything said so far.

The problem is that this working memory grows. The longer the conversation, the more notes, the more memory used. On truly long conversations, or on enormous documents fed to the model, the KV cache becomes the real bottleneck, it eats up RAM even before the model itself. Normally these scraps of paper all sit in fast and expensive memory, and when the memory runs out, the party is over.

Treating the KV cache like fuel on the SSD

And here is the idea that gives everything its title. Antirez stops treating the KV cache as something that must live only in RAM, and starts treating it like a file. He writes it to the SSD, the solid state disk, gives it a name based on a fingerprint of its content, and parks it there.

The right image is that of fuel. RAM is the engine, fast and expensive, but small. The SSD is the tank, slower but huge and cheap. Instead of demanding that all the fuel stay inside the engine, you keep it in the tank and pump in only the amount you need at that moment. DS4 uses the same logic for the model's experts too, loading them from disk only when they are actually needed, and the ones that stay switched off in a given step take up no room.

The practical consequences are quite concrete. You close the computer with a long chat open and pick it back up from the exact point, because those notes were on disk and did not vanish. You ask two questions that start from the same document and the shared part gets reused instead of recomputed. And above all, on a machine with 128 gigabytes where the model takes 76, you can work with contexts approaching a million words, something that, kept entirely in RAM, would simply be impossible.

There is a price, obviously. The disk is slower than RAM, so every now and then it waits. But precisely because the model is sparse and most experts are never touched, and because the cache can be parked without recomputing it, the trade off works out. It is real engineering, not magic.

How fast it goes and how much it drinks

Concrete numbers, because they matter. On a MacBook Pro M3 Max with 128 gigabytes, at 2 bit compression, DS4 generates around 26 words per second. When reading a long prompt, nearly twelve thousand words, it reaches hundreds of words per second. On a more loaded Mac Studio it goes even higher. These are speeds that are perfectly fine for real work, you read the answer as it comes out.

The figure that strikes most, though, is another one, the power draw. During full speed generation that laptop touches about 50 watts. It is the consumption of a working computer, not of a rack in a datacenter. A frontier model running on the energy of a laptop is something that two years ago we would have called impossible.

The point that really counts, not depending on anyone

So far we have talked about technique. Now let us talk about the reason this story matters, in our view, and it is not the speed.

Imagine you have built your product, your service, your workflow on top of an artificial intelligence model that you call over the internet from a provider. It works great, today. Tomorrow that provider can raise the price fivefold, and you either pay or shut down. It can change the terms of use. It can retire the exact model you had tuned everything around, and the next day your instructions no longer behave the same. It can decide it does not like your sector. And then there is the layer above everything, geopolitics. One model sits over here, another sits over there, and from one month to the next the rules change about what can be exported, who can use what, which services are reachable from your country. You have built a house on land that is not yours, and whoever owns the land can change the rules whenever they want.

A model that runs locally completely flips this situation. The model's weights are a file, on your disk, on your machine. No one can take it away, switch it off, limit it, or raise its price. No one sees your data, because it never leaves. If tomorrow the provider disappears, changes its flag, or gets blocked, nothing changes for you, because you already have that file. You are not betting your business on the goodwill of someone on the other side of the world.

And here is the beautiful part of the story, almost ironic. DeepSeek is a Chinese model, open, that anyone can download. Many of the most celebrated services are instead American and closed. The point is not to root for one team. The point is exactly the opposite, not to be a hostage of either one. When the weights are open and they run at your home, the game between the giants becomes their problem, not yours.

For anyone working with sensitive data this stops being philosophy and becomes practice. A law firm, a doctor, an accountant, someone who manages patents. People who cannot send that data to a cloud service, for privacy, for GDPR, for plain common sense. With a local model the data never moves, it stays on the office machine, and the AI goes to it instead of the other way around.

Where the catch is, because honesty is needed

It would be unfair to tell only the pretty side. DS4 today is not something you install with a double click. You need a machine with at least 128 gigabytes of memory, so a Mac Studio or a dedicated workstation, not your aunt's laptop. It runs a single model, by deliberate choice, it is not a Swiss army knife. It is declared beta code, so still being polished. And antirez only supports the hardware he personally owns, which is honest but means that not everything is covered.

There is also something we like, the candor about method. Antirez openly says that DS4 was written with heavy assistance from an AI model, with humans on top of the ideas, the tests, and the hunt for bugs. He does not hide it, in fact he claims it as the way software is written now. And the 2 bit compression remains a trade off, it is not free, you pay a little slice of quality, even if the structure of the model makes it far less painful than it seems.

Why I am writing about this

Because the direction is clear, and it concerns us closely. Quality artificial intelligence is ceasing to be something you can only rent from afar, and is becoming something you can own and keep under your own roof. DS4 is not yet the product for everyone, it is the proof that it can be done, today, on hardware you can buy. For a company that cares about the confidentiality of its own data, or that simply does not want to tie its future to someone else's choices, this is no longer science fiction. It is a path.

And in our view it is the right path. Not because the cloud is the enemy, but because having a choice, having control, being able to say no and still stay standing, is always better than depending.

If you want to poke around, the project is on Github as antirez ds4, and the quantized model can be found on huggingface.