The dawn of 2025 witnessed a seismic event in the artificial intelligence realm: the unveiling of DeepSeek-R1 by the Chinese team, DeepSeek. This open-source, 671 billion-parameter language model quickly established itself as a formidable contender, rivaling OpenAI’s leading models in crucial areas such as mathematics, programming, and logical reasoning. DeepSeek-R1’s ability to tackle intricate problems was particularly noteworthy, owing to its use of reinforcement learning. The model’s MIT license further disrupted the landscape by dismantling commercial barriers. The reverberations of DeepSeek-R1’s debut were felt throughout the tech world and even in the financial markets, reportedly triggering a significant downturn in AI stocks within a week of its release.
DeepSeek-R1 signified a considerable leap forward for China’s open-source AI movement in the realm of high-end language models. This unforeseen challenge has spurred global AI leaders from the United States and China to accelerate their initiatives, revealing their strategies in both technology and market positioning. This has started an AI race around the DeepSeek-R1 model.
Let’s examine how the major players in the AI arena – Meta, Google, OpenAI, Anthropic, Alibaba, and Baidu – have responded to this new competition.
Meta: Leveraging Scale and Efficiency with LLaMA 4
Meta, a frontrunner in the open-source model community, responded to DeepSeek R1 by introducing LLaMA 4. In April 2025, Meta launched LLaMA 4, its most powerful model to date, providing API access via platforms like Cloudflare. LLaMA 4 uses a Mixture-of-Experts (MoE) architecture, which divides the model into sub-models and activates only a fraction of them during each inference. This design balances large-scale parameters with inference efficiency.
The LLaMA 4 series features several sub-models, including “Scout,” with 109 billion total parameters and only 17 billion active parameters, allowing it to run on a single H100 card. The “Maverick” model has 400 billion total parameters (128 experts) but still only 17 billion active parameters, requiring a DGX cluster. This design enables LLaMA 4 to support context windows up to 10 million tokens, making it among the first open-source models to offer this capability. This is especially useful for summarizing long documents and analyzing large code repositories.
LLaMA 4 maintains rapid response times and supports multimodal inputs for images, audio, and video, thanks to its MoE architecture. Meta has chosen a strategy of efficiency, strengthening its multimodal capabilities and streamlining its operations, to solidify its position in the open-source sector while DeepSeek focuses on inference capabilities. The implementation of the MoE architecture showcases Meta’s commitment to delivering a powerful yet efficient model, accessible to a wider range of users. The long context window support further enhances its utility for complex tasks requiring extensive information processing. Meta’s open-source approach fosters collaboration and innovation within the AI community, contributing to the overall advancement of the field. By prioritizing efficiency and multimodal capabilities, Meta aims to cater to the evolving needs of developers and researchers, solidifying its position as a key player in the open-source AI landscape. The company’s strategic focus on practical applications and accessibility ensures that its models are not only powerful but also readily usable in real-world scenarios. This approach distinguishes Meta from competitors who are primarily focused on pushing the boundaries of model size and complexity.
Google: Gemini’s Evolution Towards Autonomous Intelligent Agents
Faced with the combined pressure from OpenAI and DeepSeek, Google has opted for a strategy of technological innovation. In February 2025, Google introduced the Gemini 2.0 series, featuring Flash, Pro, and Lite versions, signaling a move toward “intelligent agent” capabilities.
Gemini 2.0’s agent capabilities represent a significant advancement. The model can understand multiple modalities and actively use search engines, code sandboxes, and web browsing. Google’s Project Mariner allows AI-driven Chrome browser operations, enabling AI to fill out forms and click buttons.
Google has also introduced the Agent2Agent protocol, which allows different intelligent agents to communicate and work together, in order to support its agent ecosystem. Additionally, it has created Agent Garden, a tool and development kit to encourage third-party developers to participate.
Google is redefining the core scenarios of the next era by concentrating on intelligent agent collaboration as AI evolves towards tool-based and autonomous capabilities, as opposed to focusing on the parameter race with DeepSeek and OpenAI. The evolution of Gemini represents a strategic shift and not just a model upgrade. This shift emphasizes the potential of AI to not only process information but also to act autonomously and collaboratively. Google’s vision is to create a future where AI agents can seamlessly interact with each other and with human users, facilitating complex tasks and solving real-world problems. The Agent2Agent protocol and Agent Garden initiative demonstrate Google’s commitment to fostering an ecosystem where AI agents can thrive and contribute to a wider range of applications. By focusing on intelligent agent collaboration, Google is positioning itself at the forefront of the next generation of AI development. This strategic move allows Google to leverage its existing expertise in search, browser technology, and cloud computing to create a comprehensive platform for building and deploying AI-powered agents. The company’s emphasis on practicality and real-world applications ensures that its AI agents are not only innovative but also highly useful and adaptable to various industries and scenarios.
OpenAI: Iterating Models and Integrating Ecosystems for Reliability and Leadership
OpenAI has accelerated its model iterations and product deployments in response to DeepSeek R1. In February 2025, OpenAI launched GPT-4.5, an interim version of GPT-4, which improves logical consistency and factual accuracy, while also paving the way for GPT-5.
GPT-4.5 is considered the last major model that does not include chain-of-thought reasoning. GPT-5 will combine the features of the experimental reasoning model o3-mini and the GPT series to create a unified “general cognitive model.” OpenAI has also stated that GPT-5 will have highly adjustable intelligence levels and tool usage capabilities.
OpenAI decided to allow ChatGPT’s free users to use the basic version of GPT-5, while paid users will have access to more advanced features in order to reduce the risk of users switching to open-source alternatives. This strategy aims to keep users engaged with broad coverage.
OpenAI is also integrating capabilities like plugins, browsers, and code executors into the GPT core model, as opposed to keeping them separate, in order to create a “full-featured AI.” OpenAI is responding to the challenge of R1 by systematically integrating and increasing intelligence density. The transition towards a “general cognitive model” signifies OpenAI’s ambition to create AI systems that can reason and solve problems in a more human-like way. The integration of plugins, browsers, and code executors into the core model is a key step towards achieving this goal, enabling the AI to access and utilize a wider range of tools and information sources. By offering a free version of GPT-5 to ChatGPT users, OpenAI aims to maintain its user base and prevent defections to open-source alternatives. This strategy allows OpenAI to leverage its existing platform and user base to promote its latest model and gather valuable feedback for further development. The company’s focus on reliability and user engagement ensures that its models are not only powerful but also widely accessible and trusted. OpenAI’s commitment to innovation and continuous improvement is evident in its rapid iteration of models and its strategic integration of new capabilities.
Anthropic: Deepening Robust Intelligence with Mixed Reasoning and Thinking Budgets
Anthropic introduced Claude 3.7 Sonnet in February 2025, which focuses on “mixed reasoning” and “thinking budgets.” Users can choose “standard mode” for quick responses or enable “extended mode” for deeper, step-by-step thinking.
This method is similar to “thinking more” when people are faced with difficult tasks, as it allows AI to take longer to reason in order to improve accuracy. Anthropic also allows users to set “thinking time” to balance reasoning depth and calling costs.
Claude 3.7 outperforms its predecessor, 3.5, in challenging tasks like programming and reasoning, and is one of the few models in the industry that focuses on the transparency of the reasoning process. Its code capabilities also achieved a 70.3% accuracy rate in the most recent evaluations.
Claude 3.7 demonstrates Anthropic’s commitment to “controllable intelligence” by focusing on creating models with explainable, stable, and customizable thinking patterns, as opposed to pursuing parameter stacking. Anthropic is steadily advancing at its own pace in the R1-driven “reasoning race.” The concept of “thinking budgets” introduces a new dimension to AI development, allowing users to tailor the model’s reasoning process to specific tasks and cost constraints. This approach emphasizes the importance of efficiency and resource management, ensuring that AI systems are not only intelligent but also practical and cost-effective. Anthropic’s focus on transparency and explainability is particularly noteworthy, as it addresses the growing concern about the “black box” nature of many AI models. By making the reasoning process more transparent, Anthropic aims to build trust and confidence in its AI systems. The company’s commitment to “controllable intelligence” reflects its belief that AI should be developed in a responsible and ethical manner, with a focus on human oversight and control. Anthropic’s steady and deliberate approach to AI development allows it to carefully consider the potential implications of its technologies and to prioritize safety and reliability.
Alibaba: Building a Chinese Open-Source Ecosystem with Qwen
Alibaba’s Damo Academy quickly updated its Qwen model family just a week after DeepSeek R1 was released, releasing the Qwen 2.5 series in February 2025 and the new Qwen 3 series in late April, demonstrating strong product responsiveness and strategic vision.
The Qwen 3 series includes model versions ranging from 600 million to 235 billion parameters. It uses an MoE architecture to maintain model performance while using fewer computing resources. The flagship model, Qwen3-235B-A22B, only requires four high-performance GPUs for deployment by optimizing activation parameters, greatly lowering the barrier to entry for businesses to implement large models. In several standard tests, the overall performance of Qwen 3 exceeds that of top international models such as DeepSeek R1, OpenAI o1, and Gemini 2.5 Pro.
Alibaba places a strong emphasis on building an open-source ecosystem, in addition to technological competitiveness. Qwen 3 is fully open-sourced under the Apache 2.0 license, with open weights, training code, and deployment tools, supporting multilingual (119 languages) and multimodal applications, with the goal of creating a foundational model that can be used and customized directly by global developers.
Alibaba’s “technology + ecosystem” strategy complements DeepSeek’s lightweight breakthrough style. One emphasizes rapid iteration and leading inference, while the other emphasizes ecosystem construction and balancing scale and diversity. Qwen is gradually establishing itself as the “ecosystem hub” of open-source large models in the domestic market, a steady response to the industry disruption caused by DeepSeek. The open-sourcing of Qwen 3 under the Apache 2.0 license is a significant contribution to the global AI community, enabling developers and researchers to freely use, modify, and distribute the model. This approach fosters collaboration and innovation, accelerating the development of new AI applications and solutions. Alibaba’s commitment to supporting multilingual and multimodal applications reflects its understanding of the diverse needs of global users. The company’s emphasis on lowering the barrier to entry for businesses to implement large models is particularly important, as it enables a wider range of organizations to benefit from the power of AI. Alibaba’s strategic focus on ecosystem construction ensures that its models are not only powerful but also readily accessible and customizable, making them a valuable resource for developers and businesses alike. The company’s rapid iteration of models and its commitment to open-source principles demonstrate its dedication to advancing the field of AI and fostering a collaborative and innovative ecosystem.
Baidu: Enhancing Multimodality and Plugin Tools with the Upgrade of ERNIE Bot
Baidu significantly upgraded its flagship model, ERNIE Bot, in March, releasing ERNIE Bot 4.5 and ERNIE X1 for public testing. ERNIE X1 is positioned as a “deep thinking model,” focusing on enhancing AI’s ability to understand, plan, and execute complex tasks.
ERNIE 4.5 is Baidu’s first native multimodal large model, supporting joint modeling of text, images, audio, and video. This version also significantly reduces hallucination generation and improves code understanding and logical reasoning, surpassing GPT-4.5 levels in multiple Chinese scenario tasks.
Baidu is building an “AI tool ecosystem” that is more useful. The X1 model can use search, document Q&A, PDF reading, code execution, image recognition, web access, and business information query functions to truly realize AI’s “hands-on ability,” echoing Google Gemini’s agent route.
Baidu also announced that it will open-source some parameters of the ERNIE model by the end of June 2025 and further expand application integration with enterprise-level customers. The ERNIE series is transitioning from a closed-loop product to a platform ecosystem, attracting developers and businesses through APIs and plugin systems.
Instead of directly competing with R1 and Qwen in the open-source space, Baidu is leveraging its deep accumulation in Chinese content, search services, and knowledge graphs to deeply integrate the model with product scenarios such as search, office, and information flow, creating a more localized AI product portfolio. The development of ERNIE X1 as a “deep thinking model” signifies Baidu’s commitment to creating AI systems that can perform complex reasoning and planning tasks. The introduction of ERNIE 4.5 as a native multimodal large model is a significant step forward, enabling the AI to process and understand information from various sources, including text, images, audio, and video. Baidu’s focus on reducing hallucination generation and improving code understanding and logical reasoning reflects its dedication to building reliable and accurate AI systems. The creation of an “AI tool ecosystem” that integrates various functions such as search, document Q&A, and code execution enhances the practicality and versatility of ERNIE Bot. Baidu’s decision to open-source some parameters of the ERNIE model and expand application integration with enterprise-level customers demonstrates its commitment to fostering a collaborative and innovative AI ecosystem. By leveraging its expertise in Chinese content, search services, and knowledge graphs, Baidu is creating a localized AI product portfolio that caters to the specific needs of the Chinese market. The company’s strategic focus on product integration and ecosystem development allows it to differentiate itself from competitors and to establish a strong position in the rapidly evolving AI landscape.
In summary, DeepSeek R1’s release was more than just a technological breakthrough; it was a catalyst in the global AI arena. It has forced giants to improve inference performance, stimulated domestic companies to compete for open source, and prompted American companies to accelerate the development of agents, integration, and multimodality. The ripple effects of DeepSeek R1’s arrival have reshaped the competitive landscape, forcing companies to re-evaluate their strategies and to prioritize innovation and collaboration. The rapid pace of development in the AI field is a testament to the transformative potential of this technology and the fierce competition among leading companies.
Although the responses of the Chinese and American AI giants differ, their goals are the same: to create stronger, more reliable, and more flexible large models and win the triple competition of technology, ecosystem, and users. This process is far from over. As GPT-5, Gemini 3, Claude 4, and even DeepSeek R2 and Qwen 4 are released one after another, global AI is entering a new stage of “spiral rise.” The continuous cycle of innovation and improvement will drive the development of even more powerful and versatile AI systems in the years to come. The convergence of technology, ecosystem development, and user engagement will be crucial for success in the increasingly competitive AI market.
For enterprise users and developers, this competition will bring more choices, lower costs, and more powerful large model tools. Global AI capabilities are spreading and democratizing at an unprecedented rate, and the next decisive technological breakthrough may already be on the way. The democratization of AI will empower individuals and organizations to leverage this technology for a wide range of purposes, from automating tasks to solving complex problems. The accessibility and affordability of AI tools will fuel innovation and creativity, leading to the development of new applications and solutions that were previously unimaginable. The future of AI is bright, and the ongoing competition among leading companies will continue to drive progress and innovation in this rapidly evolving field.