In the early hours of July 17, 2026, Moonshot AI officially released its new-generation flagship open-source foundation model, Kimi K3. 2.8 trillion parameters, a 1-million-token context window, and native visual understanding — these three figures stacked together make K3 the largest open-weight model by parameter scale in the world to date. And in less than 72 hours after K3’s release, Bloomberg broke another piece of news with even greater impact on the capital markets: Moonshot AI has formally distributed shareholder resolution documents to investors and launched the preliminary procedures for a Hong Kong IPO, planning to complete its listing within six months at the earliest. At the same time, the company is advancing a new round of financing, with its post-money valuation expected to break through USD 30 billion, and some market rumors even pointing to the USD 50 billion order of magnitude. A technological breakthrough and a capital-market surge occurring in the same week is no coincidence. Moonshot AI, founded only three years ago, is using the dual-wheel drive of open-source models plus the capital markets to redefine the growth paradigm of Chinese AI companies.
Kimi K3: The Parameter Uprising of the World’s Largest Open-Source Model
In the early hours of July 17, Kimi K3 suddenly went live with no warm-up and no countdown. Moonshot AI officially defines it as “Moonshot AI’s most capable model to date,” and also the world’s first open-weight model to enter the “3T level” (trillion level). The timing of the release was precisely calibrated — it coincided with the opening of WAIC 2026 in Shanghai, forming, together with Alibaba’s subsequently released Qwen3.8-Max-Preview, a “twin heroes of Chinese open source” landscape.
In terms of architecture, K3 adopts a sparse mixture-of-experts (MoE) design, with a total parameter count of 2.8 trillion, but each inference dynamically activates only 16 of 896 experts. This approach of “brute force working miracles, yet carefully calculated” comes from Moonshot AI’s self-developed KDA (Kimi Delta Attention) hybrid linear attention mechanism and attention residual technology. According to the official technical blog, the KDA architecture boosts decoding speed by up to 6.3 times in million-token long-context scenarios, and improves overall scaling efficiency by about 2.5 times compared with the previous-generation K2, while the additional training cost is below 2%. The Stable LatentMoE routing mechanism ensures inference efficiency under ultra-large-scale parameters, avoiding the trap of traditional Dense models where “the larger the parameters, the slower the speed.”
In terms of performance, Moonshot AI maintained a rare restraint and transparency. The company officially admits that K3’s overall capability still slightly lags behind Anthropic’s Claude Fable 5 and OpenAI’s GPT-5.6 Sol, but on about 34% of benchmarks it can already compete head-on with Fable 5. Even more striking, K3 topped Arena.ai’s Frontend Code Arena evaluation with a score of 1679, taking first place in 6 of 7 front-end development categories, directly surpassing Claude Fable 5 (1631 points) and GPT-5.6 Sol (1618 points). From Kimi K2.6’s 18th place on this leaderboard to 1st place, Moonshot AI took only a single model generation.
There is also a detail that is easily overlooked: in a proof of concept, K3 ran autonomously for 48 consecutive hours, relying on open-source EDA tools to independently complete the entire process of design, iterative optimization, and verification of a 4mm² chip. This verification chip contains 1.46 million standard cells and 0.277MB of SRAM; although it still has a significant gap from a production-grade AI chip, the closed-loop verification of “AI autonomously designing an AI chip” vividly demonstrates K3’s long-horizon autonomous intelligence capability.
K3’s release even caught the attention of Musk. The Tesla CEO left a single word in the comment section of a related evaluation report: Impressive. This is already his second public endorsement of Kimi this year — in March this year, when Kimi released a paper improving ResNet residual connections, Musk had reposted it with praise. In response to Musk’s subsequent long-distance declaration that his Grok 4.6 “may surpass Kimi,” the response from Moonshot AI’s official account was rather composed: welcome to the 2-trillion-plus club.
At the commercialization level, K3’s API pricing sends a clear signal: USD 3 per million tokens for input, USD 15 per million tokens for output, and USD 0.30 per million tokens for cache hits. This pricing is in the same order of magnitude as top-tier international closed-source models, and is no longer the “cheap Chinese model” narrative. K3 has now simultaneously landed on the Kimi App, Kimi Work, Kimi Code, and the API, and the complete model weights are planned to be fully open-sourced under a Modified MIT license on platforms such as Hugging Face by July 27.
The market reaction after K3’s release was also highly dramatic. In pre-market U.S. trading on July 17, Nasdaq 100 index futures at one point fell more than 1.8%, with Nvidia leading the decline among the “Magnificent Seven” tech giants. J.P. Morgan directly called K3’s release a “DeepSeek 2.0 moment,” believing it could intensify the sell-off wave in U.S. stocks. The market’s reflex is not hard to understand: investors found that a Chinese company had achieved performance close to the top closed-source players with less computing power, and the logic of “algorithmic efficiency challenging hardware stacking” is shaking Wall Street’s faith in the return on AI’s massive capital expenditure.
Capital Surge: Valuation Up Sevenfold in Half a Year, ARR Tripling in Three Months
If K3 is Moonshot AI’s hardcore declaration at the technological level, then its moves in the capital markets are a carefully calculated strategic positioning.
According to a Bloomberg report on July 19, Moonshot AI has informed investors that the company is preparing to conduct an initial public offering in Hong Kong within six months at the earliest, and has already sent listing resolution documents to all shareholders, launching the preliminary procedures for a Hong Kong IPO. People familiar with the matter revealed that the company is currently advancing a round of financing, with a post-money valuation possibly exceeding USD 30 billion. Also, according to Bloomberg on July 22, Moonshot AI plans to launch the final round of pre-IPO negotiations in August, with a target valuation as high as USD 50 billion.
This valuation trajectory can be called “rocket-style” growth. Looking back over its financing history: after its Series C financing at the end of 2025, its valuation was about USD 4.3 billion; entering 2026, it completed three consecutive rounds of financing from January to February, in amounts of USD 500 million, USD 700 million, and USD 700 million respectively, with its valuation climbing from USD 10 billion to USD 18 billion; in May it completed about USD 2 billion in financing, with its post-money valuation breaking through USD 20 billion; in June it launched a new round of financing, with its pre-money valuation rising to USD 31.5 billion. Within half a year, the company’s valuation grew from USD 4.3 billion to USD 31.5 billion, an increase of more than sevenfold. If the pre-IPO round launched in August is completed at a USD 50 billion pre-money valuation, Moonshot AI will have achieved the leap in valuation from USD 4.3 billion to USD 50 billion in less than a year.
What supports this valuation is explosive growth on the revenue side. According to Bloomberg, Moonshot AI’s annual recurring revenue (ARR) broke through USD 100 million in early March 2026, reached USD 200 million in May, and had firmly held USD 300 million by mid-June. Caixin, citing people familiar with the matter, said that when ARR broke through USD 200 million in April, monthly revenue doubled in a single month, growing far faster than market expectations. This kind of “tripling in three months” growth rate is rare even in the global AI field.
API revenue is the core engine of this round of growth. According to Caixin, the total revenue share of the API business has already exceeded 70%, and the company has completely shed its early single model of relying on consumer-side individual subscriptions, entering a high-stickiness, high-repurchase B-side scaled monetization cycle. An even more persuasive detail: the input unit price of Kimi’s core model was raised from RMB 4 (approx. USD 0.59) per million tokens in the K2 version to RMB 6.5 (approx. USD 0.96) in K2.7 Code, a single price adjustment of 60%, yet call volume rose rather than fell — this means developers are willing to keep paying for computing power after the price increase, and the model’s productivity value has received quantified recognition on the commercial side.
The lineup of investors is equally luxurious. In addition to continued additional bets from existing shareholder Sequoia China, Alibaba, Tencent, Meituan Longzhu, China Mobile, and CPE Yuanfeng have all joined in. According to 36Kr, the company is dismantling its offshore structure in preparation for the Hong Kong IPO. Capital’s expectations for Moonshot AI’s IPO have become highly aligned — in the AI field, Moonshot AI is one of the few that can step out a listing rhythm with real money and real revenue.
Open Source + IPO: Chinese AI’s “Moonshot Moment”
Examining K3’s open-source release and its IPO sprint within the same time period, Moonshot AI’s strategic intent is clearly visible: using open-source models to establish technological influence and a developer ecosystem, and using the capital markets to establish a commercial moat and capital reserves.
At the model level, K3’s 2.8-trillion-parameter open weights directly rewrote the industry landscape. Before this, 3T-level models were almost monopolized by closed-source giants, and K3’s open sourcing means that any developer, any enterprise, can download a model close to the frontier level to their local environment and freely deploy, fine-tune, and distill it. This is not only a reinforcement of the “American closed source vs. Chinese open source” competitive landscape, but also, during the window in which global AI governance rules have not yet been fully settled, a seizing of the standard-setting power for the open-source ecosystem. Overseas media bluntly stated that K3 is breaking the ingrained perception that “American cutting-edge AI technology still clearly leads China.”
At the capital level, the timing of the Hong Kong IPO is likewise precise. In 2026, the Hong Kong Stock Exchange’s Chapter 18C (the listing mechanism for specialist technology companies) provided a listing channel for AI enterprises, and the channel for the return of Chinese concept stocks remained continuously open. Forerunners such as Zhipu and MiniMax have already obtained extremely high valuations in the Hong Kong stock market — Zhipu’s market capitalization was at one point currently about HKD 600 billion (approx. USD 76.9 billion), having peaked at HKD 1.33 trillion (approx. USD 170.5 billion), while MiniMax touched a high of HKD 410 billion (approx. USD 52.6 billion). Moonshot AI’s choice to list at this time is both a positioning move for the “New Economy 2.0” listing wave and a conversion of the primary market’s technology narrative into a secondary market’s commercial valuation.
Of course, the challenges are equally real. The costs of large-model training and inference are still oscillating at high levels, the impact of open-source models is intensifying price wars, and enterprise customers’ willingness to pay is subject to cyclical fluctuations. K3’s local deployment threshold is extremely high — according to industry estimates, its 4-bit quantized weights are about 1.4TB, and full deployment requires a super-node composed of more than 64 accelerator cards, with hardware investment approaching the level of one million USD. This means that K3’s open weights will serve cloud vendors, large enterprises, and research institutions more than individual developers. In addition, within 48 hours of K3’s release, user requests surged, and Moonshot AI announced on the evening of July 19 that it was suspending consumer-side new-user subscriptions to prioritize protecting existing paying members — this is the first case of a domestic large model proactively throttling because it was “too popular.”
But Moonshot AI’s flywheel structure has already emerged: the more users, the richer the data; the richer the data, the stronger the model; the stronger the model, the higher the paid conversion rate. When a technological breakthrough and a commercial closed loop are established simultaneously, the capital market’s premium then has underlying support.
From its founding in April 2023 to its IPO sprint in 2026, Moonshot AI took three years to travel the road that traditional tech companies take ten years to complete. The release of Kimi K3 proved that Chinese AI labs have entered the world’s first tier in parameter scale and model capability; and the preparation for the IPO marks the evolution of Chinese AI companies from “technology follower” to “commercial definer.” On July 27, K3’s complete weights will be officially open-sourced. On that day, developers around the world will have the chance to verify with their own hands just how strong this 2.8-trillion-parameter Chinese model really is. And in some trading hall of the Hong Kong Stock Exchange, the bell-ringing that belongs to Moonshot AI may already be counting down.
[Disclaimer]: The above content reflects analysis of publicly available information, expert insights, and BCC research. It does not constitute investment advice. BCC is not responsible for any losses resulting from reliance on the views expressed herein. Investors should exercise caution.