Com DeepSeek V4 Flash de 284B pot funcionar en un portàtil

A 284 billion parameter model, just ran on a laptop, not a cloud server, not a data center rack, a MacBook Pro you can buy at the Apple Store. So a guy named Anti-Rez, the creator of Redis, he built this pure C engine that fits deep seeks entire frontier model into 76 gigabytes and hits 26 tokens per second. Local AI just, yeah, it just changed. Okay, so two things happened back to back. In late April, the Chinese lab deep seek, they dropped V3 flash, and it's a 284 billion parameter model, open weights, MIT license, which I mean that means you can download it, run it, ship products on it, no permission needed. The architecture is mixture of experts, and that's the part that makes the rest possible, so like hang on to that one. So a mixture of experts model is one big model that split up into a bunch of smaller specialist subnet works. They call them experts. And for any given token, there's this routing layer that picks just a few of them to fire. So the model's got 284 billion parameters total, but only about 13 billion are actually doing work on any single token. So you get the knowledge of this huge model with the compute cost

of a small one. That's yeah, that's the trick the whole story rides on. Then within two weeks, Salvador San Felipe, right? The man who wrote Redis, he shows up with the second half, and he published DS4, also called Dwarf Star 4. It's an inference engine written in pure sea and it's built to do exactly one thing. Run deep seek, V4 flash, fully local on a machine with 1208 gigabytes of unified memory. So a MacBook Pro M3 Max, or Nvidia's DGX Spark. No cloud, No API key, no GPU cluster, none of it. And it crossed over 13,000 GitHub stars in about a month. Georgie Gorganov, the guy who wrote Lama CPP, which, you know, that's the engine half of local AI runs on, he noticed. So did the CEO of Y-Combinator, and look, when the person who built the standard tool

and the person who funds the startups, both kind of stopped to look, something real happened. Okay, so here's the problem anti-resolved. A 284 billion parameter model at full precision wants something like half a terabyte of memory. No laptop has that. So you quantize, which, that just means storing each weight in fewer bits. Drop from 16 bit numbers down to four bit, and the model shrinks by four. The catch, though, is that crushing the numbers makes the model dumber, quantize too hard,

and it starts hallucinating, looping, kind of forgetting how to follow instructions. So the usual answer is you quantize everything by the same amount and just accept the hit. Antirerez didn't do that. He went asymmetric. He sort of looked at the model and went, okay, which parts can survive being crushed and which parts can't? The expert layers, which, that's the bulk of those 284 billion parameters, those got squeezed down to two bit brutal compression.

But the parts that steer the whole thing, the routing layer that decides which experts to call, the shared expert, the projection layer, those stayed at full eight bit precision. And the logic is, the experts are redundant, right? There's a lot of them, you only use a few per token, so losing a bit of fidelity in any one of them, it's survivable. But the router has no backup. If the router gets stupid and calls the wrong experts, every downstream answer is wrong, no matter how precise the experts are. So you spend your bits where a mistake is fatal,

and you save them where it isn't. And the result fits in roughly 76 gigabytes. And on 128 gig MacBook Pro M3 Max, an independent review clocked at 26.68 tokens per second of generation. That's interactive, that's faster than you read. That's a frontier scale model answering you on your desk, on battery, wild. Now, here's the part that kind of split the community. DS4 is not a general tool. It will not run your favorite model. It won't load arbitrary GGUF files, which, you know, that's the standard format everyone uses

for local weights. It ships its own validated weight files, tested against deep sea-seeks official outputs at different context sizes, and it runs those and nothing else. And critics call that lock-in. I mean, the local AI world's already got this tangal of tools and formats, and here's a popular project just refusing to play with any of them. Fair complaint, honestly. But anti-resist argument is sharper, and yeah, I think he's right. The reason local inference has stayed a hobby for years is exactly that run anything model. You can load any model in any quance on any backend, and nobody can tell you whether the outputs any good. qualities just undefined, you're guessing.

So by owning the whole pipeline, one model, one set of validated quance, one engine, you can actually guarantee the thing works as well as it claims. And that's the difference between a toy and something you'd put in production. He picked correctness over flexibility on purpose, and it's not some thin wrapper, either. DS4 ships an HTTP API, so your apps talk to it like any other model endpoint. It's got tool calling. It's got a built-in coding agent. And it's got this disc-precisted KV cache, which that means the model's working memory of a conversation survives across sessions. So you close the laptop, come back tomorrow, the context still there. That's the kind of detail that tells you, yeah, a serious engineer built this for actual use, not for a demo.

The number that got me, it is an benchmark. It's anti-resist own line. He's been playing with local models for years, and he said this is the first time he's found himself reaching for a local model for serious work, the kind he'd normally hand to Claude. And that's the whole story in one sentence. Local models have always been the thing you use when you can't use the good one. the fallback, right? He's saying the gap closed enough that the local models now the default for real tasks on hardware he owns with zero API bills and nothing leaving the machine.

And you can kind of see the same energy in the people who piled on. A Japanese engineer spent 12 days squeezing performance out of the DGX Spark build, which that's the kind of benchmark chasing that used to need a whole data center team and a budget. On the Nvidia forums, a community contributor ported the custom quantization scheme to Kuda at the same moment anti-res was writing his own Kuda back end, just a little race in real time. metal, kuda, and an AMD rock and branch all came alive within weeks. That's not hype.

That's builders showing up because the thing works. Now, okay, the honest part, because the framing oversells in a couple places. First the hardware. Yeah, it runs on a laptop, but a 128 gig MacBook Pro M3 Max is a 3 to $4,000 machine. And the DGX Spark? That's a premium box in video markets as a personal AI supercomputer. So the democratization's real, but there's a price floor, and it sits well above what most developers have on their desk. This isn't running on your old ThinkPad, you know? Second, the quality. Crushing experts down to two bit cost you something full stop. The claim that V4 flash is quasi-frontier that comes from impressions and qualitative benchmarks,

not rigorous head-to-head of owls against the top cloud models. Reviewers flag real gotchas, edge cases in the disk KV cache, a quant trap where things kind of silently degrade. quasi-frontier is the right word. The quasi is doing some work in there, and the biggest asterisk. This is the flash model. DeepSeq's full model, V4PRO, is 1.6 trillion perimeters, and DS4's own readme says it needs very high memory machines, not a laptop. So the headline's true, a 284 billion model runs on a MacBook, but going from a usable local model to actually replacing the cloud for everything, that's still a real distance. The frontier frontier still lives in the data center. One more, too, for anyone shipping at a company.

DeepSeq is a Chinese lab. The MIT license is genuinely permissive, but the question's about training data of provenance and supply chain risk, those are unresolved. Open weights solve the privacy problem, the data never leaves your box, but they don't answer the geopolitics one. So strip the hype and here's the durable shift. Three assumptions broke all at once. That front-tier scale parameters require a cloud GPU cluster. That open weight Chinese models are the budget alternative. That local inference is a hobbyist toy. All three took a hit in the same month from one model and one engine built by one person. And for builders, the moves concrete.

If you've been paying per token API bills for Flash Tier work, you know, some orization classification, coding agent loops, internal tools, you now have a credible path to run a comparable model in-house. On one workstation, at zero marginal cost, with your data staying put. That's a real line item, and it just changed. So pick the cloud if you need the absolute top of the frontier, the 1.6 trillion parameter tier, or I don't know, you just can't justify the hardware. Like local DS4 if your work's flashed here, or your volume's high enough that the API builds sting, or privacy just isn't negotiable. That's the fork now. And a month ago it wasn't a fork at all. What to watch?

Whether the Kuda and Rockham back ends mature enough to make this real on cheaper hardware than a maxed out Mac. Because the price floor, that's the thing holding the democratization story back. And watch whether Antir ez's narrow, one-model approach or Gergonov's run everything approach wins the next year of local inference. That argument isn't settled. the question itself changed, it used to be whether local could be serious. Now it's, yeah, it's how fast it gets cheap.