Briefing
China’s Moonshot AI releases Kimi K3, the largest open-source model ever, rivaling top U.S. systems
China’s Moonshot AI announced it will release the full weights of Kimi K3 on July 27th, following its initial launch on the 16th. Packing 2.8 trillion parameters, K3 stands as the “largest open-weight model in history” as of its release date—nearly doubling the size of the previous open-weight record holder, DeepSeek (1.6 trillion parameters).
Its performance metrics are equally formidable. K3 ranked 4th out of 189 models on the Artificial Analysis composite index, putting it on par with Claude Opus 4.8 and GPT-5.5. Furthermore, it claimed the #1 spot in the Arena front-end code category, outperforming Claude Fable 5 and GPT-5.6 Sol. The gap between open-weight and top-tier proprietary models has seemingly narrowed from “generations” to mere “months.”
The market’s reaction is reminiscent of the early 2025 “DeepSeek shock.” However, a degree of skepticism is still warranted for now. Because developers cannot independently verify or modify the model until the official release, we will only find out after July 27th whether these benchmarks hold up beyond the company’s self-reported claims.
From the Perspective of Someone Who Tried K2
Grok4 vs KIMI K2 (六小虎) | 맵시’s AI 잡지식 저장소: Tistory
Exactly one year ago, I tested K2 myself. My assessment at the time was “unsatisfactory.” It couldn’t even build a proper Chrome Dino game, its Korean phrasing was awkward, its sources were heavily drawn from Chinese online communities, and despite introducing itself as “neutral,” it dodged questions regarding the Chinese government. I also suspected that the claims of it “surpassing Claude in development tasks” were likely inflated by domestic user statistics.
China’s “The Good-Enough” Strategy: What is the Real Purpose Behind It? | 맵시’s AI 잡지식 저장소: Tistory
Then this past January, I framed China’s approach as a “Good-Enough Strategy”—releasing ‘decent’ models as open source to drive mass adoption and undermine the revenue structures of U.S. tech giants.
However, K3 appears to take a step beyond this framework. Instead of aiming for just “good enough,” it claims to be right at the frontier. Its pricing also departs from the typical ultra-low-cost range of Chinese models (sub-$1) and sits closer to Anthropic’s mid-tier pricing ($3/$15). If my analysis from six months ago holds true, this signals a transition from driving adoption to full-scale monetization.
To avoid confusion, it is important to note that “open weights” and “paid APIs” are distinct product models. Releasing the weights means you can download the model file for free to run on your own hardware, whereas the $3/$15 pricing applies to the API service hosted on Moonshot’s own servers. This mirrors the structure of open-source software where the code is free but hosting is paid—a concept I previously covered regarding Z.ai’s GLM subscription case.
In the case of K3, however, this dynamic leans heavily toward the API side.
Run Kimi K3 Today: The 1.4TB VRAM Reality | Glows.ai
With K3’s weight capacity estimated at around 1.4TB, running it locally is unrealistic for individuals and most organizations. Self-hosting only becomes viable for a narrow subset of enterprises that require closed networks for data sovereignty, need to fine-tune the model with proprietary data, or have a usage volume massive enough to offset the API costs.
Ultimately, revenue will come from the API, and the practical utility of releasing the weights boils down to:
- Providing verification so users can “verify the benchmarks themselves”
- Expanding the ecosystem through hosting by third-party cloud providers
And Alibaba’s Qwen3.8
On July 19, just three days after the K3 launch and coinciding with the World Artificial Intelligence Conference (WAIC) in Shanghai, Alibaba’s Qwen team unveiled a preview of Qwen3.8-Max via X. According to the company’s announcement, it is a 2.4-trillion-parameter multimodal model. Alibaba claimed its overall performance ranks “right behind Claude Fable 5” and promised that it will also be released as an open-weight model.
However, the Max series has historically been kept as a closed-source lineup. Therefore, whether this open-weight commitment represents a genuine shift in practice or merely counter-rhetoric to offset Moonshot’s K3 announcement remains to be seen until the actual weights are released.
Another intriguing aspect lies in their corporate capital relationship. Alibaba disclosed in its 2024 annual report that it had acquired roughly a 36% stake in Moonshot AI. Although its ownership percentage may have been diluted following subsequent investment rounds, Alibaba remains a major shareholder. In essence, the competition between K3 and Qwen3.8 is akin to a race for scale taking place under the umbrella of a single backer.
Verification remains the critical hurdle. If K3’s benchmark performance is successfully replicated after the weights are released on the 27th, U.S. AI companies—which have maintained high-cost API structures—will once again face tough questions from users demanding to know exactly what they are paying for.