Open Models in 2025: China's Lead and US Response

The 2025 Open Model Landscape

• 💡 2025 marked an inflection point for open models, driven by the release of DeepSeek R1 in January and a subsequent rapid expansion of the Chinese open model ecosystem.
• 📈 Before 2025, Llama was the dominant default, with Mistral also gaining significant traction, while DeepSeek and Qwen were already producing extraordinary outputs.
• 🚀 Key releases in 2025 included Gemma and MOO 232b (the best fully open model with training data/code), followed by Llama 4 (which struggled) and Qwen 3 (which quickly gained traction).
• 🇨🇳 The summer saw a tsunami of new Chinese entrants like Minimax, Kimi, and GLM, demonstrating cutting-edge research and rapid progress.

China's Dominance and Strategy

• 📊 Qwen's cumulative adoption metrics have started to surpass Llama, becoming the new default for a long tail of tasks and users outside Silicon Valley.
• 🎯 Chinese companies, including Qwen, prioritize market share over profit and engage directly with users, releasing numerous models by small, dedicated teams.
• 💡 The Alibaba Cloud CEO articulated a vision of LLMs as the next operating system, highlighting the strategic importance of open models in China.
• 🔍 Chinese labs are doing cutting-edge research and releasing a mix of data and models, driven by community and executive pressure to adopt strong open models.

US Call to Action for Open Models

• 🇺🇸 The speaker advocates for US investment in open models to foster innovation, support research, and reduce the concentration of power in AI.
• ✅ Positive signs include Open GPTOSS (despite initial clunkiness) and a $100 million NSF grant to AI2 for making strong language models accessible to academics.
• 💰 Achieving GPT-4 level performance for fully open models, comparable to DeepSeek's 600B to trillion parameters, would require an estimated $100 million investment.

Understanding Open Model Categories

• 🧩 The speaker distinguishes between frontier models (e.g., OpenAI's GPT-4), open-weights models (e.g., DeepSeek, Qwen, Llama), and fully open-source models (e.g., AI2's MOO, Pythia).
• 🔬 Fully open-source models are crucial for researchers needing access to training data, code, and reproducibility, which is often lacking in open-weights models.
• ⚠️ Open-weights models like Llama release weights and inference code but often lack reproducible training processes or data, making them less transparent for deep research.

Future Outlook and Challenges

• 🌍 Sovereign AI is a growing concern for nations and companies, driving the need for independent compute and modeling stacks, as seen in Europe and other regions.
• 🚫 Concerns exist regarding potential manipulation by foreign entities using open models, though the speaker believes the likelihood is currently low due to structural differences from social media.
• 💡 The proliferation of local models running on personal devices could lead to significant social evolution, necessitating deeper study of their impact on businesses and daily life.