The State of Open Models

The Shifting Landscape of Open Models

• 💡 The open model ecosystem has seen a significant shift in 2025, with China emerging as a dominant force.
• 📉 Llama's previous dominance has faded, while Chinese models like DeepSeek and Qwen have gained considerable momentum and adoption.
• 🚀 The DeepSeek R1 release was a pivotal moment, prompting a recalibration in the American AI ecosystem.

China's Dominance and Strategic Approaches

• 🇨🇳 Qwen (Alibaba) is rapidly releasing a full-stack offering across various modalities (text, speech, image editing, coding) and actively engaging the community.
• 🔬 DeepSeek focuses on cutting-edge AGI research and developing models for enterprise use cases, driven by a world-class research team.
• 📈 A robust ecosystem of Chinese labs (e.g., ZAI, Moonshot AI, Tencent, ByteDance) is growing, with many exploring diverse business models and receiving government funding.
• 📊 Chinese models have surpassed US open models in performance and cumulative adoption, becoming the industry standard.

US Position and Challenges

• ⚠️ US efforts like Google's Gemma and AI2's MOO 2 32B are positive but have not yet closed the gap with leading Chinese models.
• 🇺🇸 OpenAI's GPT-OSS provides high-level cover for big tech to release open models but lacks consistent follow-through and developer engagement.
• 💰 Current US investment in open models, such as the NSF grant, is deemed insufficient to compete with frontier models, requiring a significant scale-up.
• 🕵️‍♂️ Many US startups and well-funded labs are using Chinese open-source models (e.g., Qwen) for innovation, often without public acknowledgment.

The Future of Open Models and Risks

• 🌐 Open models will persist globally, regardless of US control, as talent and resources for training diffuse across countries.
• 💻 Within two years, uncensored multimodal models (like a Sora 2 equivalent) will be runnable on personal devices, raising societal concerns.
• 🚨 Chinese models may contain propaganda or unprovable vulnerabilities, leading larger American companies to avoid them, while startups prioritize innovation.
• ⚖️ The debate between safety and openness is less urgent due to the current performance gap between frontier and open models, allowing for proactive risk assessment.

Recommendations for US Leadership

• 🎯 The US should lead in open models to align with historical values, foster innovation, and prevent the concentration of power.
• 🤝 There is a need to establish next-generation, scaled-up model centers in the US, potentially through initiatives like the American Truly Open Models (ATOM) project.
• ✅ US efforts should focus on transparency and building trust in model evaluations, leveraging academic connections, and addressing the disadvantage of not using all available data.