Introduction

On July 16, the Chinese startup Moonshot AI announced Kimi K3, an open-weight language model with 2.8 trillion parameters, which the company claims is the largest open-weight model ever released. Beyond its size, what has driven investors is the comparison: on the independent ranking by Arena.ai for front-end development, Kimi K3 has outperformed the leading closed models from OpenAI and Anthropic, offering frontier capabilities at a fraction of their price.

However, there is an important detail: the model weights will not be downloadable until July 27. For now, Moonshot has released specifications and scores, but no external researcher can yet confirm the parameter count or reproduce the benchmarks. The impressive figures for Kimi K3 remain, for the moment, those stated by the company.

Big news: Kimi-K3 by @Kimi_Moonshot is now #1 in the Frontend Code Arena with 1679 points, surpassing Claude Fable 5. This is a 17-place jump from Kimi-K2.6 (#18 -> #1). In Frontend, Kimi-K3 ranked #1 in 6 of 7 domains: Brand & Marketing, Reference-Based Design, Data & Analytics.

Rankings and Performance

In the blind ranking by Arena.ai for front-end coding, the model climbed from the seventeenth position to first place, ahead of Claude Fable 5 and GPT-5.6 Sol; in the ranking for text tasks, it is in ninth place, compared to the thirty-eighth of the previous Kimi K2.6. The available tests assign it 77.8 on Program Bench and 93.5 on GPQA-Diamond, showing an advantage over the US pair in several programming tests but a delay on more challenging software engineering tasks like FrontierSWE. On Moonshot's own internal benchmarks, K3 ranks just behind Fable 5 and GPT-5.6 Sol.

Architectural Features

Architecturally, K3 is a sparse mixture-of-experts: of the total 2.8 trillion parameters, it activates about 50 billion for each token, routing computation through 16 of its 896 experts. The context window extends to a million tokens, and the company claims a mechanism called Kimi Delta Attention that speeds up decoding on one-million-token inputs by up to 6.3 times, along with about 2.5 times better scaling efficiency compared to last year's Kimi K2.

If confirmed, the model’s dimensions would place it in a category of its own among open models. DeepSeek V4-Pro caps at 1.6 trillion parameters, and Kimi K2 itself at one trillion, while Grok 4.5 is estimated to be around 1.5 trillion: thus, K3 doubles the nearest open competitor. The model is natively multimodal on text and images, offered via the Kimi apps and an API compatible with OpenAI's SDK, and is expected to release under a modified MIT license, which Moonshot has yet to confirm in its final form.

Pricing Strategy

The sharpest weapon remains the price. Moonshot charges $0.30 per million input tokens in cache and $15 per million output tokens, beneath the rates of US operators and in line with aggressive discounts already practiced by DeepSeek. Open models generally cost on average six times less than proprietary ones, with historically slightly lower performance: it is precisely that margin that is narrowing.

Market Reactions

The announcement came amid a broad wave of selling hitting technology and semiconductor stocks. Some operators, Bloomberg reports, quickly referred to a “Kimi moment,” reminiscent of the DeepSeek shock at the beginning of 2025. The KOSPI in South Korea dropped over 6 percent, the Nikkei in Japan over 4 percent, and the Nasdaq by 1.4 percent, with Intel, Micron, AMD, and Marvell in the red.

If capable AI is becoming free, the hundreds of billions invested to build it may not pay off. Hyperscalers are heading towards $700 billion in AI infrastructure spending just this year, and Apollo economist Torsten Sloek has warned that a time lag between this spending and revenues, under pressure from the prices of Chinese and open models, could push the economy towards recession.

The question this announcement poses to closed labs was summarized by Xiaoyin Qu, a former senior product manager at Meta now an entrepreneur in AI: “When the best open-weight model surpasses the best closed-source model, how does Anthropic justify the price of Fable? Why should anyone pay for it?” Furthermore, regarding the valuation gap: “Kimi's last fundraising valued the company at $20 billion two months ago. Anthropic is worth nearly $1,000 billion, fifty times that. Why?”

The comparison uses the valuation from May: that month, Moonshot had raised $2 billion at a $20 billion valuation and is now closing a new round that would bring it around $30 billion, while dismantling its VIE structure ahead of a listing in Hong Kong.

Conclusion

A model that exists does not necessarily equate to businesses adopting it: trust, support, security, and integration remain the domain of established operators, and US labs, unlike in the dot-com era, are solidly profitable.

Chinese labs continue releasing open models that are economical and close to the frontier, and every announcement erodes the same premise: that cutting-edge AI must necessarily be expensive and American. How much Kimi K3 truly shifts the balance will be known on July 27, when the weights move from the technical sheet to something anyone can download and verify firsthand.