by Tan Aik Keong (AK)
In early 2025, the AI world was rocked by a bombshell: DeepSeek, a Chinese team behind the "deep-seeking" research effort, released its new reasoning model DeepSeek-R1 in late January. An open-source, 671-billion-parameter reasoning model, R1 matched top-tier OpenAI models on maths, coding and logical reasoning. What stood out most was how far reinforcement learning had pushed its ability to solve complex problems — and its permissive MIT licence broke down a real barrier to commercial use of frontier models. R1 didn't just dominate tech-industry chatter overnight; reports suggest it rattled capital markets too, with AI stocks reportedly taking a hit in the week after its release.
DeepSeek-R1 marked a genuine breakthrough for Chinese open-source AI at the frontier level. Facing this "small team's" surprise attack, the global AI landscape shifted fast: US tech giants and Chinese domestic players alike accelerated their response, each showing their hand on technical direction and strategy. What followed was effectively a new large-model "arms race" between US and Chinese AI companies, triggered directly by R1.
Here's how the major US and Chinese AI players — Meta, Google, OpenAI, Anthropic, Alibaba and Baidu — each responded.
Meta: LLaMA 4 counters with efficiency and scale
As the standard-bearer of the open-source camp, Meta released its next-generation model, LLaMA 4, not long after DeepSeek R1 landed. In April 2025, Meta announced LLaMA 4 as its most powerful model yet, making it available via API through platforms like Cloudflare early on. Unlike previous versions, LLaMA 4 adopts a Mixture-of-Experts (MoE) architecture — the model is made up of multiple sub-models, and only a small subset activates for any given inference, balancing massive parameter scale against inference efficiency.
The LLaMA 4 family includes several variants: "Scout," with 109 billion total parameters but only 17 billion active, runs on a single H100 GPU; "Maverick," with 400 billion total parameters across 128 experts, keeps the same 17 billion active parameters but needs a DGX cluster. This design gives LLaMA 4 an extraordinary context window — up to 10 million tokens, making it one of the first open-source models to support context at that scale, which shows up clearly in tasks like long-document summarisation and large codebase analysis.
Thanks to its MoE architecture, LLaMA 4 keeps broad knowledge coverage while staying responsive, and supports multimodal input as a foundation for image, audio and video tasks. Meta has clearly chosen a "finesse over force" strategy — while DeepSeek pushes hard on reasoning, Meta doubles down on multimodality and efficiency, securing its open-source position from a different angle.
Google: Gemini evolves toward autonomous agents
Squeezed between OpenAI and DeepSeek, Google chose to break out through a technical pivot. In February 2025, Google released the Gemini 2.0 family — Flash, Pro and Lite variants — marking its AI models' move into an "agentic" phase.
Gemini 2.0's biggest leap is agentic capability. The model doesn't just handle multimodal understanding — it can proactively call a search engine, use a code sandbox, and browse the web, genuinely acting on a user's behalf. Through Project Mariner, Google built an AI-driven Chrome browser-operation prototype, letting AI complete real interactions like filling in forms and clicking buttons.
To support this agentic ecosystem, Google also released the Agent2Agent protocol, letting different agents communicate and collaborate, with an eye toward setting the underlying standard for future AI collaboration — alongside Agent Garden tooling and a dev kit to bring in third-party developers.
As AI trends toward becoming more tool-using and autonomous, Google has stopped competing purely on parameter count against DeepSeek and OpenAI, and instead moved early to define the next era's core scenario: coordinated agentic AI. Gemini's evolution here isn't just a model upgrade — it's a strategic pivot.
OpenAI: iterating fast, integrating deep
OpenAI visibly picked up its pace on both model iteration and product strategy after DeepSeek R1 launched. In February 2025, OpenAI formally released GPT-4.5 as a transitional step beyond GPT-4, improving logical consistency and factual accuracy while laying groundwork for the GPT-5 that would follow.
GPT-4.5 is considered the last flagship model without chain-of-thought reasoning built in; GPT-5 folds in the previously experimental o3-mini reasoning model alongside the GPT series' broader capabilities, aiming for a unified "general cognitive model" architecture. OpenAI has also said GPT-5 will offer highly adjustable intelligence levels and tool-use capability.
To guard against losing users to the open-source wave, OpenAI decided free ChatGPT users would get access to a base version of GPT-5, with paid users getting more advanced features — a strategy aimed at maintaining stickiness through broad reach.
On integration, OpenAI has also stopped keeping plugins, browsing and code execution as separate add-ons, folding them directly into the core GPT model to build a genuinely "full-featured AI." Facing R1's challenge, OpenAI's answer has been systematic integration and higher intelligence density.
Anthropic: hybrid reasoning and a thinking budget
In February 2025, Anthropic released Claude 3.7 Sonnet, built around two core innovations: hybrid reasoning and a thinking budget. Users can choose a standard mode for fast responses, or switch on an extended mode that lets the model reason more deeply, step by step.
This mirrors how a person might "think a bit harder" when facing a complex task, letting the AI extend its reasoning time to improve accuracy. Anthropic also lets users configure "thinking time," balancing reasoning depth against inference cost.
Across multiple benchmarks, Claude 3.7 outperformed its 3.5 predecessor on coding and reasoning-heavy tasks, standing out as one of the few models focused on making its reasoning process genuinely transparent, with coding accuracy reaching 70.3% on recent evaluations.
Claude 3.7 reflects Anthropic's consistent focus on "controllable intelligence" — not stacking parameters for their own sake, but building a model that's explainable, stable and customisable in how it reasons. In the reasoning race R1 set off, Anthropic has kept its own, steadier pace.
Alibaba: Qwen goes fully open source
Just a week after DeepSeek R1 launched, Alibaba's DAMO Academy moved fast, releasing the Qwen 2.5 family in February 2025 and then, in a bigger move, the all-new, fully open-source Qwen 3 family at the end of April — a strong show of both responsiveness and strategic intent.
The Qwen 3 family spans models from 600 million to 235 billion parameters, using an MoE architecture to cut compute cost while holding performance. The flagship Qwen3-235B-A22B, with optimised active parameters, can be deployed on just four high-performance GPUs, meaningfully lowering the barrier to enterprise deployment. Qwen 3's overall performance has surpassed DeepSeek R1, OpenAI's o1 and Gemini 2.5 Pro on multiple standard benchmarks.
Beyond raw technical competitiveness, Alibaba is investing heavily in ecosystem-building. Qwen 3 is fully open sourced under Apache 2.0 — weights, training code and deployment tools all public — and supports 119 languages and multimodal applications, aiming to be a foundation model global developers can directly use and customise.
Alibaba's combined "technology plus ecosystem" approach complements DeepSeek's leaner, faster-breakthrough style — one emphasising rapid iteration and reasoning leadership, the other ecosystem-building at scale and breadth. Domestically, Qwen has steadily established itself as the open-source hub of China's large-model ecosystem — a measured response to the shockwaves DeepSeek sent through the industry.
Baidu: ERNIE goes multimodal, adds tool integration
Baidu carried out a major upgrade of its flagship ERNIE Bot in March, releasing ERNIE 4.5 and ERNIE X1 for public trial. ERNIE X1 is positioned as a "deep thinking model," focused on strengthening AI's ability to understand, plan and execute complex tasks.
ERNIE 4.5 is Baidu's first natively multimodal large model, jointly modelling text, image, audio and video. The new version also significantly cuts hallucination and improves code understanding and logical reasoning, surpassing GPT-4.5-level performance on several Chinese-language benchmarks.
More practically useful is the "AI tool ecosystem" Baidu is building around it. The X1 model can call search, document Q&A, PDF reading, code execution, image recognition, web browsing and business-information lookup — genuinely giving AI the ability to "do things," echoing Google Gemini's agentic direction.
Baidu has also announced plans to open-source part of ERNIE's parameters by the end of June 2025, and to expand integration with enterprise clients further. The ERNIE family is shifting from a closed product to a platform ecosystem, using APIs and plugins to draw in developers and businesses.
Rather than competing directly with R1 and Qwen on open-source turf, Baidu is leaning on its deep strength in Chinese-language content, search and knowledge graphs, integrating its models tightly into search, office and information-feed products for a more grounded, practical AI product line.
Closing: a new AI arms race, set off by R1
DeepSeek R1's release wasn't just a technical breakthrough — it was a catalyst thrown into the global AI arena. On the technical side, it forced the giants to lift their reasoning performance; on the ecosystem side, it pushed Chinese companies to race toward open source; and strategically, it pushed US companies to accelerate their work on agentic capability, integration and multimodality.
US and Chinese AI giants have responded differently, but the goal is the same: build stronger, more reliable, more flexible large models, and win on technology, ecosystem and users all at once. This process is far from over. With GPT-5, Gemini 3, Claude 4, and DeepSeek R2 and Qwen 4 all following, global AI is entering a new, spiralling-upward phase.
For enterprise users and developers, this race means more choice, lower cost and more capable large-model tools. Global AI capability is spreading and democratising at a pace we haven't seen before — and the next decisive breakthrough may already be on its way.
Part of the AK AI Corner column. Originally published in Oriental Daily (东方日报) on May 3, 2025.