DeepSeek No Longer Dominant? Meta and Nvidia Enter the Open Model Arena

After Chinese AI labs gained an early lead in the open-weight model market, Meta (META-US) and Nvidia (NVDA-US) unveiled new models this week that developers can freely download, officially entering a competitive arena led by DeepSeek, Moonshot AI, and Alibaba’s Tongyi Qianwen (Qwen). Industry insiders view this move as a clear declaration that the U.S. will not cede leadership in cutting-edge open models to Chinese firms.

Unlike popular closed models from OpenAI and Anthropic, open-weight models allow developers to download the parameters that determine an AI’s behavior, enabling inspection, modification, and deployment. However, open-weight does not always mean fully open-source, as companies may not simultaneously release training data, methodologies, or source code.

Last month, Meta, Nvidia, and other U.S. tech firms jointly urged policymakers not to impose 'premature restrictions' on open-weight models—even those originating from China. These companies argue that open-weight AI expands opportunities, fosters competition, sustains U.S. technological leadership, and prevents concentration of power among a few dominant players.

Meta and Nvidia Make Moves to Challenge Chinese Models Head-On

On Monday, Meta launched Muse Glimmer as part of its renewed strategy to reintroduce powerful models into the open ecosystem. CEO Mark Zuckerberg announced the release of weights for its latest model, Muse Spark 1.2.

Muse Spark 1.2 is considered competitive with top-tier foundation models from OpenAI and Anthropic. However, the initial models released this week by Meta and Nvidia are smaller in scale, designed primarily to run on local devices like laptops, and can be used to power on-device digital agents.

Nvidia followed up the next day with Nemotron 3.5 Lightning, continuing its Nemotron 3 model family introduced in December last year. Nvidia emphasized that its model qualifies as 'truly open-source,' as the company has released not only model weights but also the associated training datasets and training techniques for developer scrutiny.

Box (BOX-US) CEO Aaron Levie described Zuckerberg’s plan to open Muse Spark 1.2’s weights as a 'very significant event.' He noted that Meta and Nvidia’s recent actions clearly signal that the U.S. will maintain access to state-of-the-art open-source models.

Meta previously entered the open model space via Llama, but Llama 4, released in April 2025, failed to win developer favor. The company subsequently invested billions of dollars to restructure its AI division, hiring Scale AI CEO Alexandr Wang to lead the effort. Recently, Wang’s team has rolled out closed models under the Muse brand to explore new revenue streams, but now appears to be returning to the open-weight approach.

Large Government and Banking Opportunities Await—but Meta Must Rebuild Developer Trust

Meta and Nvidia still need to prove that their U.S.-based open models can attract sufficient developers and enterprise users in a market where Chinese models like DeepSeek, Moonshot AI, and Qwen are already widely embraced.

Levie believes that competitive open-weight models from U.S. firms could unlock massive demand. U.S. government agencies and major banks, concerned about security, regulation, and data sovereignty, may avoid Chinese open models but would be more willing to try products from domestic suppliers like Meta. This could allow Muse to enter sensitive application areas previously out of reach.

Charlie Dai, an analyst at Forrester, also stated that Meta’s return to open-weight models carries significant strategic weight, as it brings major U.S. frontier AI players back into the open ecosystem. For developers and enterprises, open-weight models enhance transparency, customization, deployment flexibility, and data autonomy.

However, Meta previously damaged trust among third-party developers by shifting from open to closed models. Umesh Sachdev, CEO of enterprise AI startup Uniphore, said the earlier strategy change made developers feel betrayed. Zuckerberg’s 3,500-word manifesto this week may not be enough to win them back into the Meta ecosystem.

Still, Sachdev hopes for success by U.S. homegrown players, as increased competition could lower token usage costs for AI models and accelerate innovation—ultimately benefiting consumers. Dai added that going forward, Meta must do more than just release high-performing models; it must demonstrate its ability to build a sustainable developer ecosystem.

Open-weight AI remains politically contentious in the U.S. Critics worry that Chinese models could pose national security risks and view 'distillation' techniques as potential intellectual property theft. The industry counters that distillation is a widely used method for model improvement, evaluation, and validation. With Meta and Nvidia now joining the fray, the U.S.-China AI competition is evolving from a battle over closed-model performance to a broader struggle over who controls the global open ecosystem.