Two Faces of Kimi K3: Technical Research Report Hypes It to the Sky, Professional Institutions' Actual Tests Rank It Only Second

Let's set the context first: Moonshot AI, the parent company of Kimi, has submitted a proposal to its core investors for an IPO in Hong Kong, with a potential listing within the next six months. It is currently dismantling its original VIE structure and converting it to an H-share structure more suitable for the Hong Kong market. Over this period, its valuation has risen from approximately $18 billion to over $30 billion. Against this backdrop, the newly released model Kimi K3 on July 16 can be seen as a key test for this round of IPO narrative—whether the model actually performs well directly determines whether the market is willing to pay for this valuation. In the past two weeks, several professional research reports have been released one after another. One praises K3 highly, while another offers a more 冷静 (calm) conclusion. Comparing the two together is more interesting than looking at either one alone.

I. What exactly changed in this upgrade?

K3's parameter scale reaches 2.8 trillion, making it the world's first open model to break the 3-trillion-parameter level, while supporting 1 million Token context and native visual understanding. Architecturally, it uses three new components: Kimi Delta Attention (improves attention calculation efficiency under long sequences), Attention Residuals (allows the model to selectively retrieve historical information across different network depths, rather than simply stacking them), and Stable LatentMoE (expands the number of experts to 896, activating only 16 per inference, with an activation ratio of about 1.8%). The combination of these three results in an approximate 2.5x improvement in overall scaling efficiency compared to the previous generation K2, according to official calculations. In plain terms, this means that for the same growth in parameter scale, less computational waste is required to achieve greater actual performance improvements.

A deployment detail worth highlighting is that the official recommendation is to deploy K3 using a "super-node" composed of more than 64 acceleration cards. This implies that although the model activates only a small portion of experts each time, the requirements for high-speed inter-card communication, memory management, and task scheduling are actually higher. The larger the parameter scale and the higher the sparsity, the less it saves on computation; instead, it shifts the demand from "how high single cards stack up" to "how well cards cooperate with each other." This is why this round of upgrades has made the already tight computing supply even tighter.

II. Claiming "Number One Globally," but which number one?

One widely cited claim following K3's release is that it "took the global number one spot," but this actually refers to two different things, which are often conflated. On the Frontend Code Arena leaderboard, which specifically evaluates frontend code capabilities, K3 scored 1679 and ranked first globally. However, on the broader Artificial Analysis Comprehensive Intelligence Index, K3 ranked third globally, trailing behind Claude Fable 5 and GPT-5.6 Sol, but ranking first among all open-source models. In other words, "being the strongest globally in the niche field of frontend coding" and "being the strongest globally in comprehensive intelligence" are claims of different magnitudes and should not be equated.

Specific scores for K3 on several engineering benchmarks: Program Bench score of 77.8, slightly higher than GPT-5.6 Sol's 77.6; Terminal Bench 2.1 score of 88.3, close to GPT-5.6 Sol's 88.8; FrontierSWE score of 81.2, 9.9 points higher than GPT-5.6 Sol and 14.5 points higher than Claude Opus 4.8; SWE Marathon score of 42.0, also surpassing both GPT-5.6 Sol and Claude Fable 5. These numbers themselves are fine, but a reminder is needed: most of them come from evaluation results published by Kimi itself, representing the manufacturer's self-disclosed report card, which is different from the third-party independent evaluations mentioned next. Both types of data have reference value, but neither can replace the other.

III. A more "grounded" practical research report

If leaderboard scores reflect performance on standardized test questions, a review report from a certain institution did something closer to real-world work: it tasked GPT 5.6 Sol, Kimi K3, and GLM-5.2 with completing three types of real investment research tasks—valuation modeling, multi-factor backtesting, and industry research—to see if they could truly convert abilities like long-text processing and strong reasoning into logically consistent, data-reliable, and verifiable work products. Such tests are harder to "game" and are closer to how ordinary people would actually use these models—writing reports, building spreadsheets, conducting research—rather than solving programming problems with standard answers.

Specifically, in the valuation task, Kimi K3 predicted revenue using the "volume × unit price" method. The target price, rating, and logic in the investment report were basically self-consistent, but the Excel working papers lacked effective linkage, and some formula errors affected the calculation of core data such as net profit, EPS, and DCF target prices. According to the report, "the working paper did not meet the standard for direct use." In the backtesting task, K3 retained complete Python code, stock selections for each period, portfolio returns, and net value details, showing good checkability, but failed to clearly distinguish between "signal formation day" and "trade execution day," nor did it handle transaction costs and non-tradable states like limit ups/downs or suspensions. During the test, there was even a database connection interruption. In the industry research task, the six-dimensional scoring system built by K3 was relatively complete, and the comprehensive score could be reproduced, but the Wiki knowledge base only created 2 company entity pages and 2 topic pages, with almost no source materials accumulated. In the comprehensive ranking of the three tasks, K3 ranked second, behind GPT 5.6 Sol and ahead of GLM-5.2.

To be fair, the report explicitly mentions that the testing for Kimi K3 was conducted on the day of the model's release, constituting a single test result. Sample size and prompt design may influence conclusions, requiring further verification through multiple rounds of testing. This point also echoes the phenomenon discussed in the next section.

IV. An interesting coincidence indirectly confirming the "lack of computing power" issue

Because the test happened to be scheduled on the day K3 was released, K3 experienced slower inference speeds and requests failing to respond normally during the process. The system even prompted, "Too many people chatting with Kimi, please wait and try again." The reporting party cautiously explained that this might be related to the high traffic volume and tight service resources on the first day of release, and does not represent the model's true efficiency under normal operating conditions. However, placing this detail alongside Moonshot AI's own announcement of "suspending C-end new user subscriptions," the descriptions corroborate each other. It is not that the official side is using "tight computing power" as an excuse; professional institutions indeed encountered the same problem while conducting independent tests. This counts as a rare third-party cross-validation in this round of "viral popularity due to lack of computing power" story.

V. Pricing also sends a signal

The API pricing for K3 has increased significantly compared to the previous generation: cached input price is $0.30 per million Tokens, uncached input is $3.00, and output is $15.00. This is much higher than the Kimi K2 series and also higher than most domestic models like GLM-5.2 and MiniMax.

This reflects a shift in product positioning from "low-price customer acquisition" to "flagship capability pricing driven by long context, complex reasoning, and Agent tasks." In plain understanding, when the work a model can do becomes more complex and valuable, manufacturers no longer need to rely on low prices to gain user volume. The commercial competition among domestic models is shifting from purely competing on low prices to a comprehensive contest involving model capabilities, caching efficiency, and computing power utilization rates. However, for this logic to hold, the premise is that K3's actual capabilities must support this price—which is why the more conservative practical test conclusion from the previous section is worth referencing together.

VI. What to watch next

Complete Model Weights: Planned for release before July 27, at which time more complete architecture, training, and evaluation technical reports will be available, along with more real feedback from third-party developers and enterprises, further verifying the quality of this upgrade.

Improvement in Working Paper Quality: Issues mentioned by Guojin Securities regarding Excel linkage, formula errors, and the distinction between backtest signal days and execution days are relatively specific and verifiable weaknesses. If these issues improve in subsequent versions or repeated tests of similar tasks, it indicates that engineering implementation capabilities are improving, not just the parameter scale growing.

Computing Power Expansion Progress: When C-end new user subscriptions resume and at what pace, will directly reflect whether this "card shortage" is a short-term fluctuation on the release day or a more persistent structural problem.

Hong Kong IPO Progress: The specific pace of VIE dismantling and H-share structure adjustment, as well as whether a new round of financing will land subsequently, are key windows to judge whether this round of valuation story can be realized.

Personal market observation, based on public information and verification of public research reports, does not constitute investment advice. Moonshot AI is currently a non-listed company. Information on Hong Kong IPO progress and valuation in the text comes from public reports and may change at any time; the data from the two cited research reports come from publicly released content by Huatai Securities and Guojin Securities respectively. The market has risks, and decisions require independent judgment.