Startup Thinking Machines Launches Inkling To Challenge China’s Dominance In Open-Weight AI Models

Mira Murati's AI startup Thinking Machines has unveiled its first general-purpose artificial intelligence model, marking a significant attempt by a Western company to narrow China's growing lead in the open-weight AI ecosystem.

The release comes at a time when Chinese laboratories have emerged as the dominant force in open-source AI, while several leading U.S. developers have increasingly shifted toward proprietary models.

The San Francisco-based startup, founded last year by former OpenAI Chief Technology Officer Mira Murati, introduced Inkling, an open-weight large language model that developers can download, run and customize. Unlike closed-source systems from companies such as OpenAI and Anthropic, open-weight models give enterprises, researchers and governments greater control over deployment, security, data privacy and fine-tuning.

Inkling represents Thinking Machines' first foundation model after the company debuted its AI customization platform, Tinker, in October last year. The new model is available through Tinker as well as other developer platforms.

With 975 billion parameters, Inkling ranks among the largest open-weight AI models released to date. Parameters are the mathematical variables that determine how an AI system processes information, learns patterns, and generates responses. Although parameter count alone no longer determines model quality, models of this scale typically require enormous computing resources and sophisticated training infrastructure.

The launch comes as competition in open-weight AI has shifted dramatically toward China.

Over the past year, Chinese developers, including Alibaba, DeepSeek, Baidu, and Tencent, have steadily gained market share by releasing capable open models that offer performance approaching proprietary U.S. systems at significantly lower operating costs. Their rapid progress has been aided by aggressive optimization techniques and a strategy centered on making advanced AI widely accessible rather than keeping it behind commercial APIs.

That momentum has accelerated after Meta largely retreated from its previous open-model strategy. Following the disappointing reception of Llama 4, Meta moved toward a more proprietary development approach, creating a gap among Western companies offering high-performance open models.

Thinking Machines is positioning Inkling as one of the few credible Western alternatives in that space.

The company published benchmark results comparing Inkling against proprietary models from Anthropic, Google and OpenAI, as well as leading open-weight models, many of which originate from Chinese AI laboratories. While Thinking Machines acknowledged that top proprietary systems continue to outperform Inkling overall, it said the model delivered competitive results, particularly in agent-related tasks that require AI systems to plan, reason and execute multi-step workflows.

The emphasis on AI agents reflects one of the industry's fastest-growing areas. Enterprises are increasingly deploying autonomous AI systems capable of completing complex business processes with minimal human intervention, making agent performance an important competitive differentiator.

The launch also underscores the changing economics of enterprise AI adoption.

Many businesses have begun favoring open-weight models because they eliminate recurring API costs, allow deployment inside private data centers, and reduce dependence on a single vendor. Organizations operating in regulated industries often prefer open models because sensitive information never has to leave their own infrastructure.

Thinking Machines cited Bridgewater Associates as an example of that trend. The hedge fund used the company's Tinker platform to build a customized version of Alibaba's Qwen model, which it said outperformed leading proprietary systems while operating at lower cost.

Across the AI industry, many enterprises are increasingly adapting open models to specialized use cases, improving performance while reducing inference expenses rather than relying exclusively on frontier closed models from OpenAI, Anthropic or Google.

The release also comes amid intensifying geopolitical competition over AI.

Washington has tightened export restrictions on advanced AI chips and semiconductor equipment in an effort to slow China's development of frontier AI systems. At the same time, Beijing has doubled down on open-source AI as a strategic tool to expand its global influence, particularly across emerging markets.

Chinese firms have recently accelerated the rollout of capable models and computing infrastructure, including Huawei-powered AI clusters designed to reduce reliance on Nvidia's most advanced processors. Chinese officials have also promoted open AI as a public good capable of expanding access to artificial intelligence worldwide.

Against that backdrop, Inkling represents more than just another foundation model. It signals an effort by a new generation of U.S. startups to rebuild Western competitiveness in open AI after China's rapid advances reshaped the landscape.

Whether Inkling can significantly slow China's momentum remains uncertain. However, its release broadens the pool of high-end open-weight models available to enterprises and developers at a time when demand for customizable AI systems continues to grow rapidly.