"The DeepSeek Moment, Now a Quarterly Rerun"

There's a question I end up answering for friends and family every single quarter, right on cue with the news cycle: "Has open source finally caught up with closed source?" Lately, more than ever.

Recently, AI models delivered what's being called a second "DeepSeek moment." Moonshot AI released Kimi K3 on July 16, a model with 2.8 trillion total parameters that shot straight to the top of the Frontend Code Arena leaderboard with a score of 1,679 — the first open-source model to overtake closed-source models like Claude and GPT on a frontend coding leaderboard.

The news broke, and plenty of people jumped straight into debate — it even moved AI-related US stocks briefly before the market quickly noticed a fair number of problems and pulled back.

Open Source Rising? Its Real Value Is Democratization, Not Overtaking

Anyone watching this industry has noticed that over the past year or so, this same act has repeated itself, differing only in scale — like clockwork, every quarter someone declares that some open-source model has "risen up," as if closed source's lead is about to collapse the very next second. But the reality, right up to today, is that in my own experience, the gap between open and closed hasn't actually narrowed much at all. Sure, I use both kinds of models, and open source is genuinely solid in certain scenarios — but that hasn't changed my overall view.

Let's start with the controversy around K3 itself. Just a few days after it launched, leaks claimed it had relied heavily on distillation (essentially training your own model on the outputs of someone else's advanced model) — and the director of the White House Office of Science and Technology Policy went so far as to publicly accuse K3 of distilling from Anthropic's most advanced model, Fable.

Moonshot AI naturally denied it, pointing to three proprietary techniques it says are responsible for the performance jump, framing it as original architecture. Right now it's a he-said-she-said situation, but to be fair, it clearly did make some real progress — that much shouldn't be denied.

I think the fundamental issue worth returning to is the actual real-world efficacy of open-source models. Take K3: it has a pretty basic internal contradiction — deploying at the 2.8-trillion-parameter level is still extremely expensive. You need massive server racks and top-tier GPUs just to run it. And if you need top-tier GPUs to run it, doesn't that go against the whole "open source is supposed to save everyone money" premise? It's a bit like someone telling you a restaurant is all-you-can-eat for free, but first you have to bring your own Michelin-star kitchen just to get through the door — so what exactly did you save?

Setting the controversy aside, in my own personal experience, there's still a substantial gap between the leading mainstream models and these open-source ones. Closed-source models, backed by an enormous amount of capital, are undeniably in the lead, and there's no sign of that being shaken in the near term.

I think a lot of tech enthusiasts and industry watchers are coming at this from a very different angle than enterprise decision-makers. Of course an independent developer wants models to be cheaper and easier to use — that's completely reasonable. But what enterprise decision-makers have to weigh is a much bigger set of considerations. The real value of open-source models is that they democratize the capability of models from a few months ago, or even two years ago — which genuinely helps cut costs in certain scenarios. But that's democratization, not overtaking.

As for on-premise deployment needs, cloud providers already offer hybrid cloud/on-prem options; and if what you actually need is heavy on-prem requirements and light compute requirements, an older model from well before now is probably already good enough — you don't necessarily need to chase the newest open-source flagship.

Which Model? The Real Deciding Factor Is Still Geopolitics

Then there's what I think is the single most decisive factor — and one that never shows up on any leaderboard: geopolitics.

The current US strategy is unambiguous: stay a full generation ahead in the AI race, and make sure of it. Under that logic, if your company adopts a technology stack that's geopolitically risky, you could eventually face the risk of your product being blocked from the US market altogether.

Taiwan's supply chains and listed companies' revenue structures are essentially built entirely around exporting to the US — the end buyer, at the bottom of it all, is America. So when your entire revenue structure and your entire supply chain's end customer is the United States, you're simply not going to run against geopolitical logic and adopt a model that competes head-on with American models. I think that's the fundamental logic at play here.

And model procurement decisions aren't like watching a TV drama — if a show turns out badly, you can just drop it. Once an enterprise picks a technology stack, that choice tends to be semi-permanent: the entire workflow, the data, the staff training all get locked in around it.

Every quarter has brought a fresh round of hype in the past, and model companies burn through enormous amounts of cash, giving them every incentive to maximize the marketing of their own technical edge — sometimes even concentrating their firepower purely to push up leaderboard scores. But when enterprises make technology choices, they can't afford to look only at this kind of instant, short-term advantage.

In practice, I think it's becoming increasingly clear which model companies will win over the long run — essentially the three big American players, probably plus one French company, and one to three Chinese ones. That's already plenty of qualified vendors. Most enterprises' day-to-day AI consumption costs aren't actually that high right now either — that consideration only really matters once costs climb to a meaningfully higher level.

So I'm still bullish on the three mainstream American model makers. Open-source models are obviously effective in specific scenarios, and when it comes to on-prem data deployment needs, I expect more new approaches to emerge over the next few years.

As for the "open source rises again" drama that plays out like clockwork every quarter, it's fine to watch for entertainment — but most enterprises still haven't even managed to effectively adopt AI in the first place, let alone worry about the open-versus-closed debate.