China Just Released the Largest "Open" AI Model in History. Try Downloading It.

The reaction to Moonshot AI's release of Kimi K3 this week followed a script we now know by heart. Benchmark charts. Market panic. Declarations that the frontier is no longer a closed club. TSMC fell 7 percent on the same day it reported a 77 percent jump in operating profit, which tells you the selloff was about narrative, not fundamentals. The narrative is simple and intoxicating: a 2.8 trillion parameter open-weight model, the largest ever released, now performs neck-and-neck with the best proprietary systems from Anthropic and OpenAI. Open has caught closed. The moat is gone.

Before we carve that into stone, a rude question. What exactly is open about Kimi K3?

The asterisks

Start with the most basic fact, the one buried beneath the benchmark headlines: as of this writing, you cannot download Kimi K3. The model was announced on July 16 with claims of frontier performance. The weights are scheduled to ship on July 27, eleven days later. The entire news cycle, the market reaction, the "open source catches the frontier" discourse, all of it happened during a window in which K3 was functionally a closed model with a good press release.

Then there's the price. Moonshot is charging $3 per million input tokens and $15 per million output tokens, the highest API pricing any Chinese lab has ever set. That's still cheaper than the top American models, and Moonshot deserves credit for the efficiency work that makes it possible. But the open model is not being given away. It's being sold, at premium rates for its market, through a metered API that looks exactly like the business model open-source advocates spent two years attacking.

The economics get stranger at the level of a single query. K3 currently ships with one reasoning setting, an always-on mode the company calls thinking, with a single effort level: max. Independent testers found it burning more than 13,000 reasoning tokens on a trivial image-generation prompt, roughly a quarter of a dollar for one toy query. There is no dial to turn down. A model you cannot make cheaper, quieter, or smaller is a model whose openness is conditional on the vendor's choices, which is a curious definition of open.

And finally, the elephant: 2.8 trillion parameters. Even after the weights drop on the 27th, who runs this? Serving a model of this size requires infrastructure that exists inside perhaps a few dozen organizations on Earth. For everyone else, "open weights" means the theoretical right to download a file you cannot use, and the practical reality of renting access from someone who can, whether that's Moonshot itself or a hyperscaler hosting it for you. The weights may be open, but the capability stays locked behind the same handful of doors.

What openness used to buy you

None of this is an accusation of bad faith. Moonshot published its two core architectural innovations, Kimi Delta Attention and Attention Residuals, as open research before the model launched, and the K2 line before it was genuinely usable by the community. The company has real open-source credentials.

Something else is going on. The meaning of "open" has quietly changed underneath us, and K3 is the clearest evidence yet.

In the era that made open-weight AI matter, openness bought you four concrete things: you could run the model on your own hardware, inspect and modify it, avoid vendor lock-in, and build on it without permission. Llama, Mistral's early releases, and DeepSeek's models delivered some meaningful fraction of that bundle to a wide population of developers and companies. Openness meant access.

A 2.8 trillion parameter model delivers almost none of that bundle to almost anyone. You cannot run it. Fine-tuning it is a research-lab undertaking. You will consume it through an API, which means lock-in returns through the back door. What the openness actually delivers, at this scale, is something else entirely: legitimacy, mindshare, and a geopolitical talking point. It lets a Beijing lab claim membership in a club Washington has spent three years trying to lock, timed to land days before the World Artificial Intelligence Conference in Shanghai and days after a House hearing on export control gaps. It moves markets. It recruits developers to an ecosystem. It pressures competitors' pricing.

Openness used to mean access. At three trillion parameters, it means influence.

The question hiding inside the hype

This is not a uniquely Chinese maneuver, and it would be lazy to frame it as one. Meta pioneered open weights as competitive strategy. Every lab that has ever released a model card without training data has practiced selective openness. What K3 does is push the logic to its endpoint: a model can now be simultaneously the largest "open" release in history and, for practical purposes, as closed as anything from San Francisco, because scale itself has become the enclosure.

That reframes the question everyone has been asking this week. The viral version is: who loses if open-weight models keep matching proprietary ones? The better version is: what do we actually want from openness, and does a 3T-class weight drop provide any of it?

If what we want is scrutiny, the ability for independent researchers to probe frontier systems for safety and capability, then yes, K3's weights matter enormously, and July 27 will be a genuinely important date. A frontier-class model that red teams can dissect is something the closed labs have never provided.

If what we want is competition, the picture is murkier. K3 may discipline pricing at the top of the market, but a world where frontier capability requires nation-state-scale infrastructure is concentrated regardless of what license the weights carry. The number of entities that can train or even serve such a model is the real oligopoly, and it did not grow this week.

And if what we want is access, then the most important open models of 2026 are probably not the trillion-parameter flagships at all. They're the smaller distillations and mid-size releases that actually run on hardware normal organizations own. The flagship exists to prove the lab belongs at the frontier. The small models are where openness still means what it used to.

Watch the 27th

There's a simple test coming. On July 27, either the full weights arrive as promised, unencumbered and complete, or they don't. Either the community can quantize, distill, and study this model, or it discovers the release is narrower than the announcement. Moonshot's track record suggests the weights will ship. But the fact that we spent a week treating the announcement as the event, and the artifact as a formality, says more about the state of AI discourse than any benchmark does.

The frontier may or may not be a closed club anymore. But before we celebrate the doors opening, we should check whether what's being handed out is a key or a brochure.