Kimi K3 and the Moment China Took the Open-Weight Lead

The biggest open-weight model in the world is now Chinese. Not the biggest Chinese model, the biggest one anywhere. On July 16, 2026, Moonshot AI announced Kimi K3, a 2.8 trillion parameter system it calls its most capable model and the largest open-weight release ever made. The weights are promised by July 27. You can already use it on the web and through the API.

Sit with the size for a second. 2.8 trillion parameters, downloadable, runnable on your own hardware if you have enough of it, with a license that lets you keep the weights. A year ago the working assumption in most boardrooms was that a model this good would live behind an American API, metered by the token, forever.

That assumption just broke.

Here is the shift, named plainly: the open-weight center of gravity moved east. Kimi K3 is not the whole story. It is the clearest marker yet that the frontier of models you can download, inspect, and self-host is now set in China, and that the gap to the closed American frontier has shrunk to single digits while the price has fallen by an order of magnitude.

What Kimi K3 actually is

Start with the number that matters more than the headline count. Kimi K3 has 2.8 trillion parameters in total, but it is a mixture-of-experts model, and only 16 of its 896 experts fire on any given token. That is about 1.8 percent of the network doing the work at each step. You store a very large brain and you pay to run a small slice of it per token. That is the whole economic trick of modern large models, and Moonshot pushed the sparsity harder than almost anyone shipping at this scale.

The efficiency claim is the part worth understanding, because it is where the engineering lives. Moonshot says K3 gets roughly 2.5 times the scaling efficiency of its predecessor, Kimi K2, and it credits two architectural changes. The first is Kimi Delta Attention, a hybrid linear attention scheme. The second is Attention Residuals. Both attack the same bottleneck: standard attention gets quadratically expensive as context grows, and if you want a million-token window that stays usable, you cannot pay the full quadratic bill. A hybrid linear scheme trades a little of attention’s exactness for a cost curve that grows closer to linear, so long context stops being a luxury feature and becomes the default.

The rest of the spec sheet reads like the current frontier: multimodal reasoning, a one million token context window, and an always-on thinking mode, meaning the model reasons through a problem before it answers rather than only when you prompt it to.

The benchmarks: fourth in the world, first where it counts

Numbers without a scoreboard are marketing, so go to the scoreboard. Artificial Analysis measured Kimi K3 at 57.1 on its Intelligence Index v4.1, which puts it fourth overall. Ahead of it sit Claude Fable 5 at 59.9 and GPT-5.6 Sol at 58.9. Behind it, and this is the line that will get quoted, sits Claude Opus 4.8 at 55.7.

Read that ordering carefully. An open-weight model from a Chinese lab now scores above a premium closed model from Anthropic on a respected composite benchmark. Not level with it. Above it.

The composite is the conservative reading. The specific evaluations are louder. In Arena’s Frontend Code evaluation, which puts models in blind head-to-head developer testing on real front-end tasks, Kimi K3 ranked first at 1,679 points, ahead of Fable 5. First. Not first among open models, first. And on Artificial Analysis’s long-horizon knowledge-work evaluation, the one that tests whether a model can hold a thread across a long, multi-step task instead of just answering a quiz, K3 reaches an Elo of 1547. That is up 732 points from Kimi K2.6, and it trails only Fable 5.

A 732-point Elo jump between two versions of the same lab’s model is not an incremental release. That is a lab that found something and shipped it.

“

An open-weight model you can download now scores above a premium closed model on a composite benchmark, ranks first in blind front-end coding, and trails only the leader on long-horizon work. The frontier stopped being a closed club.

”

The price is the other half of the shift

Capability alone would be a story. Capability at this price is the story. Here is what Kimi K3 costs through the API.

Set those numbers against the frontier closed models it is trading blows with on the benchmarks and the picture gets stark. You are looking at frontier-class output on several evaluations for input pricing that starts at thirty cents per million tokens on a cache hit. And that is only the hosted price. The deeper point is that the weights are coming. If the API price is too high for your workload, or you cannot send your data to someone else’s server at all, you download the model and run it yourself. There is no closed equivalent to that option, at any price.

This is the move that changes procurement. A closed model is a subscription to a capability. An open-weight model at this level is an asset you own. For a regulated bank, a hospital, a defense contractor, or any company that simply refuses to pipe its crown-jewel data through a third party, “frontier-class and self-hostable” was a wish list item with nothing on it. Kimi K3 puts something on it.

Why now: the center of gravity already moved

Here is the part that turns one impressive launch into a trend you can bet on. Kimi K3 did not arrive from nowhere. It landed on top of a shift that had already happened while most of the West was watching the closed labs.

By April 2026, according to Inference Hub, eight of the top ten Chinese AI models were open-weight, released under Apache 2.0 or MIT licenses. Not open in the vague marketing sense. Open in the download-it-and-keep-it sense. Alibaba’s Qwen family passed Meta’s Llama in cumulative downloads on Hugging Face, which means the most-downloaded open model lineage in the world is now Chinese, not American. Chinese models account for roughly 30 percent of all open-model downloads globally.

The structure underneath explains the pace. Three pure-play open-weight labs compete directly on weights, price, and release cadence: DeepSeek, Moonshot with Kimi, and Zhipu with GLM. Zhipu shipped GLM-5.2 on June 13, 2026, MIT-licensed with a one million token context, roughly a month before K3. Behind those three sit the platform giants, Alibaba with Qwen and ByteDance, who treat open models as demand generation for their clouds. When your model is a loss leader for cloud revenue, you can afford to give the weights away and price the API near cost. That is a very different incentive from a lab whose entire business is selling access to a closed model.

The demand side tells the same story from the other direction. OpenRouter’s token-routing data, which shows which models real developers actually send traffic to, tracks Chinese open-weight models going from a negligible share to a majority of tokens processed between late 2024 and mid-2026. The inflection point traces cleanly to DeepSeek-R1 in January 2025. That was the moment the price-to-capability ratio crossed the line where routing traffic to a Chinese open model stopped being an experiment and started being the default. (I wrote about that DeepSeek moment when it happened, in The Gap Closed, and about GLM’s frontier push in the GLM-5.2 breakdown. K3 is the next rung on the same ladder.)

There is the why-now. Bloomberg framed K3 as a Chinese model closing the gap with US rivals, and on everyday knowledge work the gap is now single-digit percentage points while the credit cost is often four to ten times cheaper. When the quality difference is small and the price difference is large, the price wins for most of the work most teams do. That is not a prediction. It is what the routing data already shows.

What this means for anyone choosing a model

Stop thinking of the model layer as a binary between “the good closed one” and “the cheap open one.” That framing is a year out of date. The real question now is which frontier-class model to run, and one of the strongest options is one you can hold in your hand.

If you build with AI, your option set just widened in three concrete ways.

• Self-hosting a frontier-class model is now a real line item, not a downgrade. For any workload where data cannot leave your walls, “run the weights yourself” used to mean accepting a visibly weaker model. With K3 and GLM-5.2, it means a model that trades blows with the closed frontier on public benchmarks. The compliance-driven deployment and the capability-driven one stopped being different decisions.
• Your pricing floor dropped by roughly an order of magnitude. If you are paying premium closed-model rates for tasks a K3-class model handles at thirty cents per million cache-hit input tokens, you are overpaying for the delta between 57.1 and 59.9 on tasks that never needed it. Route the bulk of the volume to the cheap frontier and reserve the premium model for the genuinely hard calls. The economics of that split just got dramatic.
• Model risk got cheaper to hedge. An open-weight model you have downloaded cannot be deprecated out from under you, rate-limited at the wrong moment, or repriced next quarter. Owning the weights is insurance against the roadmap of a company you do not control. For anything you plan to run for years, that is worth real money.

The instinct to treat the closed American labs as the permanent frontier and everything else as the budget tier is the expensive mistake here. It leads you to overpay for capability you can get for a tenth of the cost and to skip an ownership option that did not exist eighteen months ago. The teams that win the next stretch will be the ones who evaluate on capability and cost per task, decided fresh, with the open-weight frontier fully in the running. Because as of July 2026, the largest, and on several tests the strongest, downloadable model in the world was built in China, and you can keep a copy.

Sources

• Bloomberg: China’s Powerful New Moonshot AI Model Closes Gap With US Rivals
• Tom’s Hardware: Moonshot Releases 2.8-Trillion-Parameter Kimi K3
• Simon Willison: Kimi K3
• Artificial Analysis: Intelligence Index and model evaluations
• Inference Hub: Chinese Frontier Open-Source AI Models in 2026