DeepSeek V4: China’s Bid to Reshape the AI Landscape

The second breakthrough that also sparked interest is the “Engram” memory system, as announced in both arXiv and the company’s GitHub page. It bypasses the need to recompute well‑known patterns or facts, thereby enhancing the reasoning and efficiency of the LLM. A demo of the module has been released alongside the paper on GitHub, and it has been verified to increase performance levels in knowledge, reasoning, coding and mathematics.

How DeepSeek Compares to Western AI Tools

The latest available model of DeepSeek, V3.2, and other Chinese AI such as Zhipu AI’s GLM-4.7 and MiniMax’s M.2 and are currently ranked below the leaders of the pack of Claude, Gemini and ChatGPT on a pure LLM performance-based AI Leaderboards. See Figure 2 below.

With restricted access to high‑end Western AI chips, DeepSeek and other Chinese AI companies have been forced to innovate aggressively in efficiency. And by releasing many of their models as open‑source – unlike most Western models such as ChatGPT, Claude, and Gemini, which remain closed‑source – they have lowered the barriers to adoption and encouraged broader use. Their models are lean, and seem to handle the same everyday queries as their American counterparts for a fraction of the cost.

A comparison between the most popular Western AI model, ChatGPT, and the most popular Chinese AI model, DeepSeek, was conducted by running the same prompt on both platforms, providing a clear view of their comparative performance.

Prompt 1: "Find all sides of a right triangle whose perimeter is equal to 60 cm and its area is equal to 150 square cm."

Response from DeepSeek V3:

Response from ChatGPT 5.2:

Figure 3: Side‑by‑side comparison of ChatGPT and DeepSeek responses to Prompt 1

Prompt 2: "I want to plan a week of tasty vegetarian dinners that are high in protein, take under 30 minutes to cook, use mostly overlapping ingredients to reduce waste, and fit a moderate budget. Can you suggest a meal plan, a combined shopping list, and also adjust one of the meals to be nut-free for an allergy"

Response from DeepSeek V3:

Response from ChatGPT 5.2:

Figure 4: Side‑by‑side comparison of ChatGPT and DeepSeek responses to Prompt 2

Following the recent breakthroughs, DeepSeek V4 has the potential to “dominate the ‘autonomous coding’ sector, moving beyond simple snippet generation to manage entire software repositories with human-level reasoning.” This could put direct pressure on GitHub Copilot, Cursor, and other Western coding assistants that have dominated the Western market until now.

A Tale of Two Markets: Western Resistance vs. Global Adoption

Although DeepSeek’s R1 model became highly popular following its release last year, its usage in Western countries has remained limited. In Australia, its adoption is reported to be minimal. Concerns about data security and censorship have led to the AI being banned from government devices. Outside the government, cybersecurity company Netskope claims that almost 70% of its “Australian customers have enacted blanket bans” on the application.

However, in countries outside the West, it has been growing in popularity. Microsoft released a study on AI usage globally, and found that DeepSeek has demonstrated “strong dominance in developing countries”, due to its open-source releases and high cost-effectiveness. It has reached countries and regions that have not been able to develop the AI infrastructure needed for power-intensive LLMs.

Since August 2025, Chinese AI has exceeded Western AI in terms of cumulative downloads in the open-source community, indicating the increasing popularity of Chinese AI. Being relatively more open-source than most Western models has also meant that DeepSeek and other Chinese AI are increasingly being used as the preferred foundation for building AI solutions.

Why Australia Should Pay Attention

If Australia intends to develop its own sovereign AI industry, it cannot rely solely on the example set by American companies such as OpenAI and Anthropic. Chinese AI scientists are now reaching breakthroughs in LLM efficiency and low-power computing that could be far more relevant to Australia’s economic needs, energy limitations and environmental priorities.

While many American AI companies are continually pushing for ever‑larger, compute‑hungry models, equivalent Chinese counterparts are seeking ways to further reduce the need for compute power, easing the financial and environmental burden. With Australia’s data‑centre electricity consumption projected to triple to 6% of total grid demand by 2030 – against the backdrop of coal retirements and mounting grid instability – energy‑efficient AI architectures are emerging as the more economically appropriate and sustainable path forward.

Concerns about Chinese AI – particularly around issues like censorship – are legitimate and understandable. However, that does not mean that AI such as DeepSeek should be completely cut off from the Australian AI industry. The breakthroughs that Chinese AI engineers are achieving at startups such as DeepSeek, MiniMax, Moonshot AI and Zhipu AI, combined with the relatively more open‑source availability of their models, represent a valuable resource for Australia’s sovereign AI ambitions. Ignoring these advancements would mean missing out on capabilities that could meaningfully shape Australia’s AI future.