Silicon Valley is quietly building its next generation of

Harvey, the $11 billion legal AI unicorn, released Tenet, a specialized legal model built on Kimi K3, achieving nearly double the success rate on its proprietary LAB benchmark compared to the raw base model. The move reflects a shift from API renting to model ownership, with US startups like Cursor, Cosine, and Devin also using Chinese open-weight models such as Kimi K2.5, K2.6, and K2.7 as foundations for their products.

Silicon Valley is quietly building its next generation of These failures highlight a massive gap: general-purpose models are great at answering questions, but they are terrible at executing professional workflows. A lawyer doesn't need a chatbot; they need an agent that can navigate thousands of pages, identify risks, and draft a precise memorandum. This is exactly why Harvey—the $11 billion legal AI unicorn—just released Tenet, a specialized model built specifically for complex legal tasks. While many assume a company of that scale would only use OpenAI, Tenet actually uses Kimi K3 as its base foundation. Why "Base Models" aren't enough for professional workflows If you are a developer building an AI agent /en/tags/ai%20agent/ , you know that prompt engineering has its limits. You can write a massive system prompt, but you can't easily teach a model the "intuition" required to know when to fetch an additional document or how to spot a subtle discrepancy in a contract. Harvey's decision to use an open-weight base like Kimi allows them to perform deep post-training. By training on specialized legal datasets, Tenet achieved nearly double the success rate on Harvey’s proprietary LAB benchmark compared to the raw base model. This reveals a massive shift in the AI industry. We are moving away from the "API Renting" era and into the "Model Ownership" era. The Old Way API Renting : You call Claude /en/tags/claude/ or GPT-4 via API. It's fast to start, but as you scale, your margins vanish into OpenAI's pocket. You own nothing. The New Way Foundational Refinement : You take a high-quality open-weight model like Kimi or DeepSeek /en/tags/deepseek/ , perform supervised fine-tuning SFT with your proprietary industry data, and deploy it. You own the intelligence. The hidden "Chinese heart" in US AI products It turns out that many of the most famous "American" AI tools are actually powered by Chinese foundations. The technical community is seeing a pattern where top-tier startups skip the trillion-parameter training race and instead "import" high-performing Chinese models to use as their starting point. The massive AI coding tool recently admitted that its Composer 2 model is built on top of Kimi K2.5. Cursor /en/tags/cursor/ : Cosine: The specialist tool for maintaining legacy enterprise code uses Kimi K2.6 to power its Lumen Outpost model. Devin: The much-hyped "AI Software Engineer" revealed that its SWE-1.7 release is based on Kimi K2.7. This isn't just about being cheap; it's about performance and control. By using these models as a "pre-fab" foundation, these companies can focus their compute and engineering talent on the top 5% of the problem—the industry-specific logic—rather than trying to teach a model how to speak English or write basic Python from scratch. Even Mira Murati’s new venture, Thinking Machines, is playing in this sandbox. While they are training their own "Inkling" model, they are leveraging Kimi K2.5 to generate the synthetic data needed for their supervised fine-tuning. We are witnessing the rise of a new AI supply chain. Chinese models are becoming the "industrial raw materials" for the next wave of specialized AI agents in the West. They aren't just competing for users anymore; they are becoming the invisible architecture upon which the next generation of Silicon Valley giants is being built. Why AI coding assistants are actually making our technical debt 15h ago /en/news/7612/ Why human kids are still way more efficient at learning language 1d ago /en/news/7518/ AI coding tools are turning into a dopamine trap for developers 1d ago /en/news/7448/ Built a shared brain for AI agents after markdown file chaos 3d ago /en/news/7235/ EU copyright office confirms AI output falls outside protection 4d ago /en/news/7124/ Team messaging that actually remembers why you built that feature 4d ago /en/news/7106/ Next Taican 700 is officially here and it’s bringing NOA to the → /en/news/7696/