Following the open-source release of Kimi K3, the open-sourcing of top-tier models has become a realistic possibility.

Following the open-sourcing of Kimi K3, the industry has witnessed the world's first 3T-parameter model with fully open weights, which is currently the largest open-source AI model in the world by parameter scale.

Over the past three years, a key shift in the global large AI model sector is that Chinese model vendors have gained a dominant edge in the open-source community ecosystem to a greater extent than their US counterparts.

K3 further amplifies this effect. Kimi stated in its official introduction that although the overall performance of K3 still lags behind the strongest closed-source models Claude Fable 5 and GPT-5.6 Sol, it demonstrates cutting-edge capabilities in its full set of evaluations and steadily outperforms all other models.

Compared with Kimi K2, K3 has a 167% increase in total parameters, a 220% rise in activated parameters, and the context window is expanded from 128K to 1M, an 8-fold improvement. Theoretically, its computational complexity will grow at a squared level.

The core reason why K3 can push the parameter count to 2.8 trillion and achieve a 2.5x improvement in scaling efficiency (higher model capability under the same computational volume) lies in two underlying optimizations to the traditional Transformer architecture: KDA (Kimi Delta Attention / Mixed Linear Attention Mechanism) and Attention Residuals.

The former can compress the ever-growing KV Cache into a fixed-size matrix state, while the latter can effectively reduce the information attenuation and gradient vanishing problems in model training, ensuring smooth information flow.

Meanwhile, K3 adopts technologies including Stable LatentMoE (Stable Latent Mixture of Experts System), MoonEP (underlying communication library), and FlashKDA (high-performance computing operator) to significantly optimize inference computing cost, communication capability and computing efficiency.

After landing on Hugging Face, K3 topped the trending list with more than 4,000 likes within 30 minutes. Hugging Face CEO Clem Delangue posted that this is the fastest release growth rate ever recorded.

OpenAI President Greg Brockman talked about K3 in a recent interview, saying that it is an undoubtedly competitive new AI model, and revealed based on relevant assessments that the gap between China and the United States in model development may be only about 4 months.

In addition, Elon Musk has also publicly praised the research of the "Attention Residuals" paper underlying the K3 architecture as "impressive".

01

Apart from the influence brought by open-sourcing itself, an AI large model researcher told a reporter from Jiemian News that the more direct impact of K3's open-sourcing lies in the commercial dimension. Combined with lower prices and the gradually narrowing model performance gap, Chinese open-source models have fundamentally undermined the dominance of closed-source model vendors such as OpenAI and Anthropic over API pricing power, which involves a huge amount of potential commercial interests behind it.

Before the launch of K3, another sensational closed-source model was Anthropic's Fable 5. There is no officially announced exact parameter count for this model, but the industry generally speculates that it is a 5T-10T level model.

The skyrocketing popularity of both Fable 5 and K3 has, to varying degrees, dispelled the long-standing doubts in the industry: does it still make sense to continue increasing model parameters?

The aforementioned researcher told Jiemian News that model vendors have almost no other choice but to continue "Scaling Up" if they want to stay in the game. "At least for now, post-training techniques such as RL (Reinforcement Learning) mainly serve as 'embellishments'. The better the base model is, the higher the starting point of post-training will be."

Chinese model vendors have been successively following up on the R&D of models with parameters exceeding 2T. Alibaba released the preview version of Qwen3.8 Max on July 19, a multimodal model with a parameter scale as high as 2.4 trillion; in early July, reports said that MiniMax is developing a new generation of large language model with 2.7 trillion parameters, internally referred to as "M3 Pro".

02

It is reported that both models will be open-sourced immediately after launch. In fact, this has become an important way for Chinese model vendors to expand their technological influence.

Wenxin 5.0, officially launched by Baidu in January this year, is a 2.4T-parameter model according to its official introduction, which may be the first domestic vendor to reach this scale, but this model has not aroused much discussion at home and abroad.

On the one hand, this model adopts an ultra-sparse Mixture of Experts architecture, with the proportion of activated parameters less than 3%, which is around 70B. In comparison, K3 has around 104B activated parameters. To some extent, the size of activated parameters will also affect model performance.

On the other hand, although Wenxin 4.5 joined the open-source ecosystem last September, Wenxin 5.0 is still a closed-source model. Some industry insiders pointed out that this is a collaborative route of open-sourcing base models and keeping flagship models closed, which is intended to strike a balance between inclusive technology that attracts users and commercial benefits generated by the closed flagship model.

However, China's open-source model ecosystem is growing stronger. The closed-source nature of Wenxin 5.0 means that a large number of developers cannot deploy it on their own and conduct more technical analysis by observing and studying its internal structure, thus losing the voice from the geek group to a certain extent, which has become increasingly important for the current influence of models.

For catching up with larger-parameter models, apart from a series of comprehensive challenges at the software and hardware levels, the more intuitive dilemma facing vendors is the shortage of computing power.

According to analysis by industry insiders, for deployment alone, K3 requires one 8-card B300 server, while its training requires a 10,000-card cluster of top-tier AI GPUs. Just the day after its launch, Kimi urgently released a notice of computing power shortage and suspended new membership sign-ups.

In this regard, startups once again show disadvantages compared with large tech giants. A person from an intelligent computing center said that most of its current computing power has been locked by large key clients such as top-tier tech giants. For startups represented by Kimi, this means their purchasing power lags behind, and it is difficult to supplement computing power in time even when facing scenarios like the explosive demand for K3.

Another cloud vendor insider revealed the composition of its clients: the top two clients are top-tier hardware giants and internet giants respectively, and the total monthly revenue from all other model vendors combined is less than that of either of the two. Moreover, due to the scarcity of top-tier GPUs, cloud vendors are also facing the dilemma of delayed delivery of client orders, and only 10% of the first client's order of thousands of servers was delivered in the first half of the year.

03

Even so, the goal of model vendors to develop better models and open-source them is, to some extent, urgent.

In almost the same week when Kimi K3 was open-sourced, Jensen Huang, together with Microsoft, Meta, IBM, Intel and 22 other tech giants, published an open letter to support the open-source ecosystem, and OpenAI and Google soon signed to join it.

Probably under the dual pressure of public opinion and commercial interests, Dario Amodei, CEO of Anthropic, who has always held a firm stance on technological blockade, also recently publicly stated that Anthropic has never advocated banning open-weight models.

With more influential calls, the global trend of model open-sourcing has become increasingly unstoppable. From the perspective of inclusive technology, this is undoubtedly positive, but for enterprises, the all-open-source scenario will make them more transparent to users from technology to cost, and it will not be easier for them to achieve breakthroughs and survive than it is today.

This article is from "Jiemian News", reporter: WU Yangyu, editor: WEN Shuqi, published with authorization from 36Kr.