Friday, July 17, 2026

Kimi K3, 2.8 trillion parameters, open source. This thing got a score of 1026 on Hacker News, which means that it is recognized abroad. Dark Side of the Moon was quite ruthless this time, directly throwing out the world's first open source 3 trillion level model. Moreover, it was not that kind of PPT open source, but it could really run. The architecture uses a mixture of linear attention and attention residuals, which to put it bluntly is to find a new balance between cost and effect. What does this mean for developers? This means that you finally don't have to kneel and ask for APIs. For the 2.8T model, you can pull down and adjust it yourself, do vertical domain or privatization deployment, and take off directly. But don't be happy too soon. With the scale of the model parameters, you can't run without an A100 cluster. Ordinary developers can only use APIs to play with them. The main problem in engineering lies in memory and reasoning speed. Officials say it supports 1 million token contexts, but when deployed, memory bandwidth and latency will make you doubt life. In terms of commercialization, Dark Side of the Moon is obviously a move to grab the niche: open source attracts developers and then use the enterprise version to make money. But the question comes. Now that the open source model is rolled up like this, with Llama, Qwen, and DeepSeek running all over the place, can K3 really kill it with parameter levels and architectural advantages? My judgment is: there is a breakthrough in technology, but implementation still depends on ecological supporting facilities. If the supporting tool chain and community are done well, a wave of corporate customers can be harvested; otherwise, it will be another tragedy of "I open source, but you can't use it".

Talk about the mobile phone market again. Realme has escaped, and OnePlus has also suspended operations in Europe and the United States. These two signals together show that China's mobile phones have entered a cruel knockout round when going out to sea. In the past, realme relied on "daring to leapfrog" to improve cost-effectiveness. As a result, now it cannot even defend the domestic market and can only retract overseas. One plus is even worse. It directly stops Europe and the United States, and the OPPO system is shrinking. What's behind this? It is the AI entrance war that is forcing mobile phone manufacturers to take sides again. You see, Apple's AI has just been registered and connected to Qianwen. Huawei Xiaoyi, Xiaomi Surge, and vivo Blue Heart are all madly involved on the end-side AI. Mobile phone manufacturers now not only have to sell hardware, but also have to raise large model teams. Who can afford this cost? Sub-brands such as realme and OnePlus have neither scale advantages nor AI capabilities, so they are inevitable to be abandoned. Anthropic's analysis put it well, don't make hardware, just bet on models, because "the models are strong enough, there is no need to occupy the entrance." But the reality is that if mobile phone manufacturers want to survive, they must differentiate AI, otherwise it will be the next realme. For ordinary users, the good thing is that end-side AI is finally about to be implemented on a large scale. After Apple's AI entered China, Siri can finally work seriously; the bad thing is that there are fewer and fewer mobile phone brands, and there may be only a few giants left in the future. You like to buy it or not.

Finally, I will say something more heart-warming. The daily salary of DeepSeek interns is 5500, and the number of doctoral students is expanded to 200,000 per year, but there are only 35,000 college positions. Put these two news together, it is the sense of division in the AI industry. Top AI talents are being robbed of money, but the doctors at the bottom cannot even find teaching posts in universities. This is actually a systemic problem: what the AI industry needs is people who can build wheels, train models, and implement projects, not doctors who simply publish papers. There are too many doctors, but people who can really work are still scarce. I suggest that friends who want to study for a PhD should first ask themselves: Can you compete for a job with DeepSeek interns after graduation? If you can't, switch to engineering as soon as possible and don't stick to academic circles.

The core thread of today's technology signals points to two intertwined waves: one is the accelerated iteration of generative AI and the deepening of the "model-as-service" model, and the other is the development of technology giants and emerging forces in hardware, software, and market strategies. Differentiation and game.

The release of the KIMI K3 model, especially its parameter scale of 2.8T and its performance approaching the top level in evaluation, marks another solid step in catching up with technology for domestic large models. This is not an isolated incident. Previously, we saw that Anthropic chose to "light hardware and heavy models" and focus on providing powerful basic model capabilities. Its products such as Claude Code and Cowork are all trying to empower work scenarios through the models themselves. In China, the dynamics of Tencent's Hunyuan Hy3 and WeChat WeLM also show the giant's two-line layout at the AI model and application levels. These signals collectively point to a trend: the boundaries of models 'capabilities are constantly being broadened, and the "Model as a Service"(MaaS) business model will become more mature. In the next six months to a year, we may see the emergence of more customized models in vertical areas, as well as the explosive growth of APIs and application ecosystems around the output of model capabilities. This will make AI no longer the exclusive preserve of a few technology companies, but an extension of "computing power as a service" that can be reached in all walks of life.

At the same time, competition for hardware entrances and adjustments in market strategies are also quietly taking place. realme's withdrawal from the China market, focusing resources overseas, and OnePlus's suspension of operations in the US and European markets all reflect the fierce competition in the smartphone market and the painful choices of brand strategy. This is in sharp contrast to the previous successful registration of end-side AI models by brands such as Apple, Huawei, OPPO, and vivo. As AI capabilities gradually sink into terminal equipment, hardware differentiation will become particularly important. Although the fluctuations in SpaceX's share price after its listing are not directly related to AI, they reflect the capital market's caution about the valuation logic of high-tech companies, especially when its business model has not yet been fully verified. In the future, we may see intensified competition among mobile phone manufacturers in AI computing power and applications, and AI capabilities will become an important label distinguishing the high-end and mid-to-low-end markets. Companies with core technologies and unique application scenarios on the AI model side, even if they are not involved in hardware, may become important players in the ecosystem through APIs or collaborative models, just as Anthropic is betting.

Another phenomenon worthy of attention is the structural contradiction between the "high fever" and the "surplus of doctors" in the AI talent market. The daily salary of AI internships is 5500 yuan, which on the one hand highlights the extremely high demand and talent premium in the AI field. On the other hand, the surge in the number of doctoral graduates in colleges and universities and the lag in job supply expose the problem of the disconnect between higher education and industrial demand. This indicates that the cultivation and employment of AI talents will face greater challenges in the future. High-end talents will still be scarce, but non-top academic talents may face a shortage of students. This structural contradiction may prompt more people to turn to AI application development, Agent construction and implementation, and niche areas where AI is integrated with specific industries, rather than just pursuing breakthroughs in the model itself.

Overall, the underlying technology of AI is iterating at an unprecedented rate and rapidly penetrating into various application levels. Competition for hardware entrances remains fierce, but model capabilities themselves are becoming increasingly important carriers of value. The structural imbalance in the talent market indicates that the AI ecosystem will shift from technology-driven to a stage where it focuses more on application implementation and talent matching. In the coming year, we will see more "model-as-a-service" innovations and how companies and individuals can adjust their strategies to adapt to this rapidly changing world amid the AI wave.

KIMI K3 claimed to have 2.8 trillion parameters and 1 million token contexts, and was immediately packaged as an "open source 3 trillion level model." Technically, mixing linear attention and residual mechanisms can indeed increase operator density on specific tasks, but the actual effect is still limited to official benchmarks and is far from reaching GPT‑5.6 or Fable‑5 of the same scale. More importantly, the training and reasoning costs of such models are still on the order of billions of dollars, which only a few cloud vendors can afford. Even if open source is open, deployment thresholds and energy consumption are daunting for most companies, and the real commercial value is packaged as "cutting-edge scientific research." If there are no supporting ecological tools and cost control solutions in the future, K3 is likely to become a "show-off" rather than a productivity.

At the same time, the recruitment craze in the AI industry was ignited by the news of interns earning 5500 yuan a day, but it also reflected the other side of the imbalance between supply and demand of talents. High salaries can only attract a few top talents, but they do not have the ability to simultaneously improve project management, data governance and security compliance. Correspondingly, there is a sharp expansion in the supply of doctoral doctors: since the expansion of enrollment in 2020, more than 100,000 new doctoral doctors will be added every year in 2025, while university positions are shrinking. A large number of highly educated talents are forced to flow to non-scientific research positions or become inefficient "AI armchair". Historically, the computer hardware bubble of the 1970s and the Internet bubble of 2000 have experienced a similar cycle of "talent surplus-technological optimism-capital pursuit", which eventually led to the failure of a large number of projects due to the lack of sustainable business models.

The capital level is also not optimistic. SpaceX's market value fell by one-third in the first month, and short sellers 'book profits were nearly US$4 billion, indicating that the market's tolerance for high valuations has dropped to a critical point. The financing boom of the AI model is in sharp contrast to the IPO failure of traditional hardware companies: the former relies on "technical black boxes" to attract capital, while the latter is forced to cut prices or withdraw due to unclear commercialization paths. The chase of capital often pays the price first. The real risk lies in that when the marginal improvement in technical performance is blocked by cost, energy consumption and regulatory barriers, the profit margin of the entire ecological chain will be compressed, leaving only a few companies that can control and make breakthroughs in compliance governance. Taken together, the current technical gimmicks, talent structure and capital valuations all reveal signs of a bubble. Short-term heat may continue to push up valuations, but sustainability in the medium and long term still needs to be carefully evaluated.