Three days ago, I met a long-time law firm partner over coffee. He shared with me, "Since AI became mainstream, I haven't needed to hire new assistants for a while. AI can handle most entry-level tasks. If I do hire someone in the future, they'll need to excel in areas where AI falls short." I agreed and added, "They should also be adept at leveraging AI more effectively than the average person—that's a significant competitive edge."

My friend primarily uses general-purpose large models like Claude and GPT, noting, "There are numerous plugins tailored for legal work—I need to explore them." But what was the outcome? He confessed, "They're still a bit cumbersome to use and can only address minor issues. They're like snacks, not a full meal." He didn't anticipate this issue being resolved anytime soon.

This conversation reminded me of a recent news story: U.S. legal tech company Harvey unveiled its proprietary foundational model, Harvey Tenet, designed specifically for legal professionals. Although I'm not in the legal field, it seems my lawyer friend's predicament might soon have a solution.

Harvey stands as the world's most valuable legal AI company and one of the hottest AI vertical application unicorns, with a valuation hovering around $11 billion. OpenAI is also among its investors. It serves over 1,300 institutions and more than 100,000 lawyers across 60 countries, with over half of the Am Law 100 firms as its clients. Legal services represent one of the highest-value, lowest-tolerance-for-error industries globally, demanding far more from AI applications than consumer-grade or even many enterprise-grade solutions.

Harvey Tenet, a foundational model targeting professional legal clients, was built on Kimi K3 and trained with legal expertise data over two months using about 150 B300 units in collaboration with Fireworks AI. According to Harvey's internal assessment, it achieves roughly twice the task completion rate of the base model in legal tasks. Currently, Tenet is only available as a research preview and hasn't been fully integrated into applications—its real-world performance remains to be seen.

This immediately brought to mind Cursor (recently acquired by SpaceX for $60 billion), the world's leading AI coding application, which released its foundational model Composer 2 five months ago. With exceptional coding capabilities and low inference costs, it quickly garnered widespread attention. The world was curious: Where did this model originate? How was it trained? Later, Cursor revealed it was commercially licensed and post-trained based on Kimi K2.5. The entire AI community was taken aback!

In just half a year, first Cursor, then Harvey—one in coding (the most AI-intensive, token-consuming field globally) and the other in legal services (the highest-barrier, most professionally demanding service industry). These aren't isolated incidents but two confirmations of the same trend: Open-weight models + industry expert data + post-training = Any vertical company can refine its own cutting-edge model! I refer to this as the 'Prometheus Moment' for global AI applications.

Specifically, for highly specialized vertical applications, the previous mainstream approach relied on closed-source cutting-edge models with various expert plugins—as my lawyer friend did. While this offers the most advanced standardized AI solutions, the drawbacks are evident: First, non-specialization—you're utilizing a 'general model,' and plugins are built on top of it, not optimized for niche fields. Second, high cost—everyone is aware of the exorbitant API access fees for cutting-edge closed-source models. Third, lack of openness and flexibility—adjusting the model or deploying it locally is nearly impossible.

Harvey Tenet and the earlier Cursor Composer 2 validate a new approach: With such advanced open-weight models available, why not rely on the more cost-effective and adaptable open-source ecosystem? The drawbacks of closed-source models are precisely the strengths of open-source ecosystems. To illustrate: From an application standpoint, 'closed-source models + plugins' are akin to standardized fast food from a central kitchen—meeting most people's baseline needs but rarely excelling in specialized verticals. 'Open-weight models + industry expert data + post-training' are like custom dishes tailored to target customers' tastes—high-value meals at reasonable prices. My lawyer friend's complaint about 'snacks vs. main courses' is thus resolved.

Globally, the most active and community-driven open-weight models predominantly originate from China. Both Cursor and Harvey chose Kimi as the foundation for their proprietary models—a result of careful technical evaluation, more a necessity than a coincidence. Without exaggeration, Chinese open-source models like Kimi have become the 'common foundation' for top global AI application companies.

Other domestic model vendors have also achieved notable success. For instance, Harvey announced three additional expert model research projects: two (M&A due diligence, review forms) based on Zhipu GLM-5.2 and one (law firm knowledge) on Alibaba Qwen 3.8. All four projects utilize Chinese open-source models as their foundation! This competitive landscape reflects China's overall AI research prowess. Without domestic models, the path of 'open-weight models + industry expert data + post-training' wouldn't exist. From this perspective, Chinese vendors are propelling this trend forward.

Media outlets like the Financial Times list Kimi and DeepSeek as Chinese vendors 'winning clients from Silicon Valley and Europe,' putting pressure on Anthropic and OpenAI. Especially in providing foundational models for vertical applications, Kimi already boasts two highly successful cases—and this is just the beginning. If this approach works for coding (the highest AI token demand) and legal services (the lowest error tolerance), it can be replicated in finance, healthcare, manufacturing, and other verticals. That's why I refer to this as the 'Prometheus Moment' for AI applications, with Chinese open-source vendors like Kimi playing the role of Prometheus.

Cursor and Harvey are foreign vertical application companies, but I believe the domestic market holds equal, if not greater, potential. Having studied the Chinese software industry for eight years, I'm aware of the enormous scale of enterprise applications here—and traditional software hasn't adequately addressed customer needs (the reasons are complex and won't be elaborated here). Importantly, most unmet enterprise application needs can be addressed through AI! Many are already working on this. I believe the domestic industry AI application market will soon explode, possibly becoming the world's largest.

This also provides a clearer answer to the commercialization of AI open-source ecosystems. Many still question whether open-weight models have viable business models—after all, 'open weights' imply anyone can download and deploy them locally, suggesting a low commercialization ceiling. However, it turns out that a vertical application ecosystem based on open-weight models has almost limitless commercial potential and is certain to establish a mature and stable business model. The release of Harvey Tenet may be a small step for Harvey and Kimi but a symbolic leap for the entire AI open-source ecosystem.

This doesn't mean closed-source models like Claude and GPT are obsolete. I'm a paying user of multiple cutting-edge closed-source models—many issues still rely on them for now. Note that while launching its proprietary model, Harvey still integrates Claude Opus 5, GPT-5.6 Sol, and other frontier models, announcing new integrations as recently as July. Tenet is an addition, not a replacement. Meanwhile, Anthropic and OpenAI's annual recurring revenue (ARR) continues to grow rapidly, potentially approaching or exceeding $100 billion by year-end. This is good news for the industry: The market is large and growing fast, allowing mainstream vendors to profit—a healthy, not overly competitive, industry! I believe the generative AI commercialization wave is just beginning, with open-source ecosystems gaining market share and closed-source models maintaining core strengths. Everyone has a bright future.

One undeniable fact: The surge in open-weight models has exerted immense pressure on cutting-edge closed-source models, exemplified by GPT-5.6 Luna's dramatic price cuts—making it a 'best value' candidate in many eyes—and Claude Opus 5's lower pricing compared to Fable 5, along with the cancellation of Sonnet 5's price hike plans. Without pricing pressure from open-source competitors, such concessions from closed-source models would be unimaginable.

Competition from open-source camps benefits the world—all users enjoy better token value! This, I believe, is the greatest contribution of Chinese open-weight vendors like Kimi to the AI industry. Everything has just begun, and I eagerly await more contributions.