Meta to unwind $2b Manus AI deal on China order

Chinese AI dominance under threat as Meta launches open-source models for US market

Meta Platforms has entered the open-weight artificial intelligence (AI) arena with renewed vigor, introducing the Muse Glimmer model, which can run locally on personal computers, and announcing plans to release the weights of its latest flagship Muse Spark 1.2 model. This move comes amid concerns over potential US regulatory restrictions on Chinese AI models.

CEO Mark Zuckerberg announced in a blog post that Meta intends to resume releasing open-source models to deliver personal superintelligence to billions of people and small businesses. Analysts suggest that American users and companies may prefer to build with American models to avoid compliance, reputational, and other risks associated with Chinese AI models.

Meta's new offerings are positioned to directly compete with top Chinese alternatives across both the flagship and lightweight tiers. The company's Muse Spark 1.2 model, released last week, sits between Anthropic's Claude Sonnet 5 and OpenAI's GPT-5.6 Terra in overall capability, while the lightweight Muse Glimmer model effectively matches China's 1-trillion-parameter Kimi K2.5.

These developments come as Washington debates the rise of cost-efficient Chinese open-weight models, leading US frontier labs to warn of dangerous shifts in global AI leadership. While no restrictions have been enacted yet, Meta believes anxieties about compliance will push enterprise customers away from Chinese options.

The company's move is seen as a response to the growing adoption of Chinese models among global enterprise customers seeking to cut AI expenditure. Meta's lighter models offer strong intelligence for their compact footprint, effectively matching China's top-tier models with the added advantage of being "not Chinese." Industry experts believe that more US open-model providers could make Chinese models less attractive due to regulatory uncertainty. However, Meta's success will ultimately depend on model performance.