Moonshot AI released Kimi K3 on July 16, a 2.8-trillion-parameter model billed as the largest open-weight system ever shipped. At launch it lands third on the Artificial Analysis leaderboard, behind Claude Fable 5 and GPT-5.6 Sol, yet it beats every rival on front-end code. China just cut its distance to the US labs down to a matter of weeks.
Key Takeaways
- Kimi K3 runs 2.8 trillion parameters in open weights, with full weights due to ship on July 27
- It tops the Frontend Code Arena with a 76% preference rate against Claude Fable 5, at $15 per million tokens versus $50
- Its release lined up with Xi Jinping’s World AI Conference speech and a sell-off in chip stocks on Wall Street
Have an AI Sum Up This Article
ChatGPTWhat Moonshot Actually Shipped
Kimi K3 is a mixture-of-experts model. Its architecture fires only 16 of its 896 experts per token, roughly 1.8% of the network on each pass, which is how it runs 2.8 trillion parameters without lighting up the whole thing per request. Moonshot describes it as the first open system in the 3-trillion class.
The model carries a one-million-token context window and native vision, handling text and images in the same exchange. Moonshot laid out these specs in the technical brief posted on its official site, ahead of the full weight release.
One word does the heavy lifting: open. Moonshot plans to publish K3’s weights on July 27, which will let developers download, inspect and run the model themselves. Until then it stays reachable only through the lab’s own interface and API.
The rankings paint two different pictures. On the general Artificial Analysis index, K3 comes out third, behind Fable 5 and GPT-5.6 Sol. On Arena, it takes first place on the front-end code bench with a 76% preference rate in blind testing, and it posts 88.3 on Terminal-Bench 2.1. On web and mobile, it offers three reasoning effort levels, Standard, High and Max.
This profile echoes a path already opened on the Chinese side, back when Zhipu shipped a free open-access rival to Claude with GLM-5.2. Kimi K3 pushes the same logic a step further, in size and in raw performance on one specific task.
User Impact: Front-End Code First
For developers, the sharpest signal comes from Arena. In blind testing, developers preferred Kimi over every leading US model on front-end web development, including Fable 5 and GPT-5.6 Sol. The strength does not cover all of coding, but it hits a daily use case for product teams.
Price backs the message. K3 charges $15 per million output tokens, about three times less than the $50 of Fable 5. It stays pricier than other Chinese models, GLM-5.2 at $4.40 and DeepSeek V4 at $0.87, but it lands well under the equivalent US frontier models.
Open weights change the nature of the offer. A team will be able to host K3 on its own infrastructure, audit it and adapt it, without depending on a vendor’s server. That point separates a closed model from one a company controls end to end, an argument other Chinese labs already lean on, like DeepSeek raising $1.5B as it preps an IPO.
The catch sits in the calendar. Until the weights ship, nobody can independently verify the model or run it locally. July 27 will separate the promise from real availability, as often happens with model announcements.
More articles on Horizon
- Anthropic Cuts Fable 5 Limits Starting July 20
- GPT-5.6 Deletes User Files in Full Access Mode
- AI Overviews Now Fall Under German Media Law
Rival Impact: Wall Street and the US-China Race
K3’s release lined up with Xi Jinping’s speech at the World AI Conference in Shanghai. The markets noticed the timing: the Nasdaq slipped about 1% on Friday, with investors trimming chip names like Nvidia. A frontier-level open model out of China hits the thesis of a durable US lead head on.
Pressure moves onto price. With Fable 5 at $50 and an open model three times cheaper right behind it on a visible task, the US labs lose part of their pricing room. The differentiator shrinks to the few-week gap measured by the general leaderboards.
The debate over the model’s real reach stays open. Dean Ball, who leads strategy work at OpenAI, admitted Kimi was a very good model whose performance cannot be explained away by simple distillation. Other voices pushed back, pointing to the lack of dangerous cyber capabilities and to China’s own incentives to eventually rein in its open models.
Behind the model, Moonshot is moving fast. The lab raised about $2 billion at a valuation near $20 billion, and is preparing a Hong Kong listing. Kimi K3 serves as its public proof point, at a moment when the question is no longer whether China is catching up, but by how many weeks.
Our read stays cautious until July 27. A front-end code ranking does not make a universal model, and the weight release will deliver the real usage verdict. The rest of the race plays out there, on what developers actually run in-house.
Follow the story on Horizon.
