Western AI Startups Turn to Chinese Open-Weight Models for Core Systems

Western artificial intelligence developers are increasingly adopting Chinese open-weight models as foundational layers for specialized software, according to Chinese state media reporting. Startups including US legal-tech firm Harvey, code-editor platform Cursor, Cognition, and the UK's Cosine have built production tools atop Moonshot AI's Kimi models using domain-specific post-training and reinforcement learning. Furthermore, Mira Murati's Thinking Machines Lab adopted DeepSeek-V3's mixture-of-experts architecture and utilized Kimi K2.5 for synthetic data generation when developing its Inkling model, underscoring how Chinese open architectures are influencing global vertical AI stacks.

This dynamic marks a structural shift in global AI development pipelines. Rather than competing solely at the consumer application layer, Chinese open-weight releases are operating as base models, architectural templates, and synthetic data engines for Western software startups seeking cost-effective alternatives to closed-source proprietary APIs.

Overseas artificial intelligence developers are increasingly moving beyond proprietary closed-source ecosystems to build specialized commercial tools on top of Chinese open-weight models, according to a report by People's Daily. Rather than functioning solely as end-user conversational interfaces, systems such as Moonshot AI's Kimi and DeepSeek's architectures are serving as foundational bases for specialized post-training, architectural blueprints, and sources of synthetic data. In the legal sector, US artificial intelligence company Harvey recently unveiled Tenet, its first post-trained open-weight model. The system was developed using Moonshot AI's Kimi K3 foundation. Harvey disclosed that Tenet was trained over roughly two months across approximately 1,750 specialized legal agent environments. Earlier testing conducted by Harvey in May indicated that most general-purpose foundation models completed fewer than 10% of practical legal tasks under strict evaluation criteria. Following specialized post-training, Tenet nearly doubled the number of fully completed complex legal tasks relative to the base Kimi K3, while increasing completion rates in drafting, review, and contract negotiations by approximately 20%. The adoption trend extends to software engineering platforms. The US company behind the Cursor code editor disclosed in technical documentation that its Composer 2 and Composer 2.5 models were built upon Kimi K2.5. Cursor refined the base model through continuous pre-training on code-centric datasets alongside large-scale reinforcement learning to develop specialized agent capabilities. Similarly, US startup Cognition and British AI startup Cosine have utilized Kimi models as base systems, applying reinforcement learning to strengthen software engineering performance in their primary products. Chinese open-weight systems are also guiding model design and training methodology abroad. Thinking Machines Lab, an AI research venture founded by former OpenAI Chief Technology Officer Mira Murati, disclosed in technical notes for its open-weight model Inkling that its mixture-of-experts architecture primarily follows the technical blueprint established by DeepSeek-V3. The lab also reported using open-weight models, including Kimi K2.5, to generate synthetic data for its initial round of supervised fine-tuning. As the broader industry shifts from simple conversational chatbots toward agentic systems capable of executing complex multi-step workflows, Chinese open-weight systems are emerging as key structural components across the international AI software stack.