Yuval Avidani
Author
China's Moonshot AI just open-sourced Kimi K3 - 2.8 trillion parameters, the biggest open model the world has ever seen, and anyone can download it starting July 27th. And that's not even the surprising part. The surprising part is that a monster this size is even runnable at all, and the answer hides in three letters: MoE.
Let's break this down, because underneath that huge number sits the real lesson of 2026 - how to build a powerful model without access to the newest chips.
2.8 Trillion Parameters - And Why That's Even Possible
Parameters are the numbers a model learns during training, and roughly speaking, more parameters means more "room" to learn in. But here's the catch: a dense model with 2.8 trillion parameters would need to light up every single parameter every time it generates a word - and that's expensive to the point of impossible. This is where the trick comes in.
K3 is built as MoE, Mixture of Experts: instead of one giant brain, there are 896 small "experts," and for each token only 16 of them light up. Think of it like a hospital: it has hundreds of specialist doctors, but when we show up with a sore throat, they don't call in the entire staff - they call an ENT doctor and maybe one more. All the expertise exists, but only the relevant slice works at any given moment.
The result: out of 2.8 trillion parameters, only a tiny fraction - about 50 billion, less than 2% of the model - actually computes at each step. That's what lets a model this size run fast and at a sane cost.
The Numbers: Leading the Open Models, Breathing Down the Closed Ones' Neck
Let's be precise. K3 isn't "the strongest model in the world" - it's the strongest among the open ones, and surprisingly close to the closed frontier. On Artificial Analysis's intelligence index it scored 57, the highest score any open-weights model has ever achieved. In the Frontend Code Arena, where developers rate code blind, K3 beat out Claude Fable 5 and took first place.
The most exciting part is code and terminal work. On Terminal-Bench 2.1, a benchmark for autonomous terminal operation, K3 scored 88.3% - just half a point behind GPT-5.6 Sol (88.8%), one of the strongest closed models in the world. It also leads on long-horizon coding benchmarks like SWE Marathon and Program Bench.
It's also important to say the other side: several sources point out that K3's hallucination risk wasn't published with full transparency, and on certain automation tasks it's still lagging. A strong open model is great, but independent verification matters even more when the maker chooses what to publish.
The Second Bit of Magic: 4-Bit Quantization
There's a second innovation here worth understanding, because it's the thing that makes downloading this monster possible at all. Usually you train a model at high precision (16 bits per number), and only afterward "shrink" it down to lower precision to save memory - a process called quantization, which usually costs you a bit of quality.
Moonshot flipped that: they trained K3 from the start at 4-bit precision (a format called MXFP4), so the model learned to compensate for the compression during training itself. The result is impressive: the weights file weighs about 1.4 terabytes, instead of the roughly 5.6 terabytes it would need at regular precision. It's still not running on your laptop - you need a cluster of several servers - but it's a massive leap in accessibility, and the format is natively supported on Nvidia's Blackwell chips and AMD's MI400.
Why This Is a Geopolitics Story, Not Just an AI One
And here's where the bigger picture comes in. The US has been restricting the export of advanced chips to China for years now, on the assumption that without the newest hardware, it's hard to train frontier models. K3 is proof that efficiency can make up for hardware. Instead of brute force, Moonshot invested in smart architecture: hybrid linear attention, an optimizer at the attention-head level, and quantization-aware training - which together, by their own reporting, deliver 2.5x the training efficiency of the previous generation, Kimi K2.
In my view, this is the part that's easiest to miss under the "biggest model" headline. The bottom line isn't the size - it's the proof that you can reach the frontier with fewer chips, if you work smarter. And for anyone building on open source, that's excellent news: one more strong, open player, spreading the power around instead of concentrating it in two or three closed companies.
Bottom Line
In my view, Kimi K3 is two stories in one body. The first is technical: MoE and 4-bit quantization are the reason a 2.8-trillion-parameter model is even runnable at all - it's not the size, it's how much of it works at any given moment. The second is geopolitical: China is showing that efficiency beats export restrictions. My reservation is a fair one - not every number was published transparently, and you need a serious cluster to run this thing - but the direction is clear.
So if the biggest open model in the world is already breathing down the neck of the closed frontier, and it comes from a company that's supposed to be hardware-limited in the first place - what does that say about the closed giants' edge?