Transcription
There is a specific moment I keep coming back to. April 24th, 2026. Someone posts DeepSeek V4 Pro on Hugging Face at midnight, and within a few hours, developers are running sweet benchmark evaluators on it. By morning, the numbers are everywhere, and they don't make sense because the numbers are too good at a price that sounds way too good to be true.
That's the thing about DeepSeek, they keep doing this. They release something, people run the numbers, and the reaction is always the same. Wait, how? So, today I actually want to answer that. Not just that DeepSeek V4 Pro is a good model, but what did they actually do differently, and why does it matter that they released it as an open-source model?
If you follow DeepSeek's research output over the past 2 years, I'm sure you've noticed the pattern. Every paper is fundamentally about the same thing, doing more with less. Less compute, less memory, bandwidth, training costs, everything is efficient, and they're very strategic about this. Chinese labs operate under export controls that limit their access to high-end GPUs. So, DeepSeek had a specific incentive that Western labs didn't. Make the architecture more efficient because you can't just buy your way out of the problem.
The biggest achievement with V4 Pro is that the model can now go up to a million token context window. This is basically how much the model can hold text in its memory at once. But, the important part to understand is that having a million token context window doesn't mean anything if running the model costs you a fortune. And with the standard way models process text, a million context window would be really expensive. Roughly speaking, every time you double the context, you quadruple the compute time. That means at a million tokens with old-style processing, you'd be burning money so fast, it just wouldn't be viable for most teams to continue. The million token context window would be possible, it just would be practically useless.
So, what did DeepSeek do? How did they pull off a million token context window? They did it by changing how the model reads. Let's say someone drops a 500-page report on your desk and says you have a meeting in an hour where you're going to present the main parts of this report. And no, assume ChatGPT doesn't exist. So, of course, you won't be reading every page equally. You'll probably skim through the summary, figure out which sections are most relevant to what you actually need, and then go deeper on those areas. You still looked at the whole thing. You just didn't burn equal energy on every page. That's basically what DeepSeek built. We're not treating every word as equally important and comparing it against every other word, which is where compute cost usually goes crazy. It now groups chunks of text, makes summaries of those groups, and then only goes in on the parts that seem relevant to your question. So, you're still reading everything. You're just not spending equal energy on every single word. The result shows at a million tokens of context, V4 Pro uses 10% of memory their previous model used. And their previous model was already considered optimized.
Now, the other parts of this is why is this cheaper than other models? V4 Pro technically has 1.6 trillion parameters, which in theory sounds really expensive. But the secret is that you never actually use all of them at once. Let's do a quick analogy. There are hundreds of specialists and staff members at a hospital. When a patient comes in, nobody pulls every doctor into the room. You route the patient to the two or three people who are the most relevant. DeepSeek runs a similar concept. For any given text, V4 Pro only fires up about 49 billion of those 1.6 trillion parameters. The rest are sitting there idle. So, you get the depth of a huge model without paying to run a large model on every single word.
DeepSeek's API cost is about $3.48 per million output tokens. Compare that to Claude Opus 4.7, it's $25. So, if you're a solo developer, maybe you don't care, but if you're running millions of AI calls a day, that cost difference changes what's viable for you to build. Also, if you self-host the open weights, each call is basically zero. So, you've got a model with a million token context window running on a fraction of the memory at a fraction of the cost. That alone is brilliant, but cheaper and faster means absolutely nothing if the outputs are garbage. So, the real question is, does it work?
But, before that, back to the video in just a second. I've been using Nex AI and it's been saving me hours every week. There's three things that I use it for. First, slides. Anytime I'm preparing for a meeting or doing research for a video, I used to spend hours. Now, I just paste my outline and something like build the slide deck on the future of AI agents, professional tone, dark theme, and I've got a fully structured presentation in a few minutes. Second, their file generator agent. I basically typed create a content calendar template for an AI video creator with weekly topics, formats, and upload dates, and it gave me a fully organized, ready-to-use spreadsheet with everything drafted out for me. You just add your own content to make it personal. And most importantly, I use it as my daily assistant. I told the agent builder, "Gather the most important AI news every week and email them to me on Monday." It asked a couple of quick follow-ups, and that was it. Agent created. The best part is that it connects to the tools you're already using, Google Calendar, Slack, Confluence, and so many other ones. It seamlessly plugs into your workflow. Click the link in the description to start your free 7-day trial, and there's no credit card or anything required. You can use my code on the screen for 10% off the pro plan. Now, back to the video.
So, Deep Seek did something smart with training. We are no longer throwing data at one giant model and hoping it figures everything out. They instead train separate specialist model, a coding model, a math model, a writing model, each one focused on their own domain expertise. Each one rewarded for actually getting the right answer. For example, does the code run? Does it pass the test? For those of you familiar with this architecture, this is mixture of experts or MOEs. Then they merged all these specialist models into one final model. And honestly, the final model has really good results. On Swee benchmark where it gives the model a GitHub issue and asks it to fix the bug, V4-Pro scores 80.6%. If we compare that to Claude Opus 4.6, it's essentially identical. On competitive programming, V4-Pro has the highest rating of any model currently available.
Now, is it perfect at everything? No. Independent testing found it's about 6 to 8 months behind on broad general intelligence. So, things like reasoning, knowledge, open-ended stuff, DeepSeek even says this themselves in their own paper. But on coding, which is what most companies are actually paying for right now, it's leading the way. And this is no accident. They picked the domain with the most immediate economic value and went all in on coding.
Now, here's what I think is the most interesting part of this whole thing. DeepSeek didn't do this alone. In the span of 12 days in April, there were four separate Chinese AI labs that released frontier coding models. They're all open weight, free to download, and they all drop within the same week. Now, remember, export controls for Nvidia chips were supposed to slow this down, basically limit the hardware and preserve the American lead. And in the narrow sense, they did slow it down. These labs don't have the same access to high-end GPUs that labs in the US do. But it's impressive that the hardware pressure didn't stop them. It just forced them to get really, really good at building efficient models. All the clever architecture we just talked about, the group detention, the specialist training, none of that came from having more resources. It came from having fewer. And now all of this is sitting on hugging face for free.
Is there still a gap? Yes, it lives in the safety, the general reasoning and task where there's no clear right or wrong answer. But on software engineering, there is little to no gap at all. That's what makes V4 Pro worth paying attention to. It's not about V4 Pro is a good model, but how it got here. Less hardware, smarter architecture, open weights built under pressure, and basically released to everyone. It's the people's model. Follow and I'll keep you posted.