DeepSeek went nuclear this week โ€” launching V4, its most powerful open-source model to date, and simultaneously adding AI vision capabilities for the first time, closing what had been a glaring gap against US rivals. The Hangzhou-based start-up’s one-two punch reverberated through Chinese chip markets, sent analysts scrambling to reassess stocks from Cambricon to SMIC, and drew a direct response from Beijing: President Xi Jinping called for “disruptive innovation” at a symposium on basic research. Meanwhile, the week also delivered a landmark Chinese court ruling on AI job displacement and a new push for AI-powered judicial efficiency in Shenzhen. Here is this week’s China AI Weekly.

DeepSeek V4: The Whale That Now Sees

DeepSeek released its V4 series this week, describing it as the most powerful open-source platform yet โ€” capable of challenging OpenAI and Anthropic on their own turf. V4 delivers top-tier performance in coding benchmarks and shows significant advantages in reasoning and agentic tasks, according to the company. (Source: SCMP)

On April 29, DeepSeek followed the V4 launch by adding multimodal capabilities for the first time โ€” meaning its chatbot can now process images and video in addition to text. Senior researcher Chen Deli announced on social media that “the little whale can now see,” a reference to DeepSeek’s whale logo. The new “image recognition mode” joins the “expert” and “flash” modes introduced earlier this month, bringing the start-up in line with rivals that already offer multimodal support. (Source: SCMP)

Why It Matters for Enterprise Teams

The V4 launch lands with two implications beyond the benchmark scores. (For the full CIO decision framework covering Entity List risk, Azure deployment, model-agnostic architecture, and governance, see our DeepSeek V4 Enterprise Decision Framework.)

First, the China chip ecosystem gets a demand jolt. DeepSeek V4 is highly compatible with domestically made chips, and analysts at BOC International say this will accelerate the commercialisation of AI computing power in China. Su Lingyao (BOC International) noted that V4 has “lowered the threshold for using high-performance AI models” and will offer “more affordable AI capabilities to small and medium-sized enterprises or even individuals.”

Potential beneficiaries include chipmakers Cambricon Technologies, Moore Threads Technology, Hygon Information Technology, and MetaX Integrated Circuits, as well as fab operators SMIC and Hua Hong Semiconductor. The China AI chip market is projected to grow from 142.5 billion yuan in 2024 to 1.34 trillion yuan ($196.2 billion) by 2029 โ€” a 54% compound annual growth rate, according to Guotai Haitong Securities. (Source: SCMP)

For enterprise architects planning inference hardware roadmaps, the trend line is clear: Chinese-built AI chip capacity is accelerating, and vendor lock-in considerations now include a geographically distinct supply chain option. Any inference strategy that does not account for this bifurcation is incomplete.

Second, the vision gap is closed. DeepSeek’s lack of multimodal support had been seen as a competitive weakness since its breakout moment in January 2025. With V4’s vision capabilities now in beta, teams evaluating DeepSeek for production use cases that require combined text and image understanding โ€” document processing, content moderation, visual analytics โ€” can now include it alongside Qwen-VL and other Chinese multimodal options.

Hong Kong’s Sovereign AI Push: HKGAI-V3 on DeepSeek V4

The Hong Kong Generative AI R&D Centre (HKGAI) announced plans to unveil its HKGAI-V3 model in the first half of 2026, built on DeepSeek V4 architecture with “full-parameter fine tuning for localisation.” The model is optimized to run on both mainstream hardware and domestic chips, including Huawei’s Ascend 910C.

HKGAI director Guo Yike indicated the lab is exploring overseas expansion of its “sovereign AI” capabilities โ€” a term defined in HKGAI’s research as the ability for a jurisdiction to develop and operate AI systems on its own infrastructure, ensuring “security of critical technologies, the mastery of data sovereignty, and alignment with local language, culture and institutional requirements.” (Source: SCMP)

This is the first concrete signal of Chinese-influenced sovereign AI models being exported to markets outside mainland China. For multinational enterprises operating across multiple jurisdictions, the emergence of region-specific AI stacks โ€” DeepSeek-based, Huawei-chip-optimized, Hong Kong-localized โ€” adds a new variable to model procurement decisions.

Xi Jinping Calls for “Disruptive Innovation” in Basic Research

President Xi Jinping addressed a symposium on basic research in Shanghai on April 30, urging China to pursue “original and disruptive innovation” amid an escalating US tech race. He called for expanding China’s talent pool and fostering an innovation culture that is “open, inclusive and tolerant of failure.”

China invested nearly 280 billion yuan ($41 billion) in basic research in 2025, accounting for 7.08% of total R&D expenditure. Xi called for gradually increasing that proportion and creating a “diversified investment landscape.” (Source: SCMP)

The timing โ€” days after DeepSeek V4’s launch โ€” is not coincidental. The message signals continued state-level support for compute-intensive basic research, which underpins the long-term trajectory of Chinese AI labs and their ability to sustain rapid release cycles.

A Hangzhou Intermediate People’s Court ruled it illegal for a company to terminate an employee on the grounds that an AI replacement would be cheaper โ€” a landmark decision that affirms limits on AI-driven job displacement.

The case involved a 35-year-old worker surnamed Zhou who oversaw AI-generated responses at a fintech firm. The company told Zhou his role could be replaced by AI. The court disagreed, awarding him over 260,000 yuan ($38,067) in compensation.

Judge Shi Guoqiang told state broadcaster CCTV: “We don’t believe AI technology has reached the point where it can substantially replace human workers.” A Guangzhou court reached the same conclusion in a 2024 case involving a graphic designer. (Source: SCMP)

For multinational teams managing AI adoption in China, this ruling carries practical weight. AI automation plans should account for Chinese labor law constraints โ€” replacing headcount purely on cost grounds carries legal exposure that is now explicitly on the record.

Shenzhen’s AI Judges: 50% More Cases Processed

Shenzhen’s Intermediate People’s Court reported that judges processed 50% more cases in 2025 with the help of an AI-powered judicial assistance system โ€” the first in China to use domain-specific large language models. Each judge handled an average of 744 cases, up 249 from 2024, making Shenzhen’s judges the most efficient in Guangdong province. The system now covers 85 judicial procedures across civil, administrative, and criminal litigation, and will be rolled out to dozens more Chinese cities. (Source: SCMP)

This is one of the more concrete public-sector AI ROI figures to emerge from China. The system was self-built by the court, not procured from a vendor โ€” a pattern that suggests China’s judicial AI push may follow a custom-built, city-by-city deployment model rather than platform licensing at scale.

Also This Week

  • Hangzhou deployed humanoid traffic police robots on its streets โ€” China’s latest visible step toward real-world embodied AI applications. (Source: SCMP)
  • Song Yuhang, the Oxford AI chip star who co-founded Fractile in the UK, has returned to China to join Nanjing University as an associate professor after departing Fractile in May 2024. (Source: SCMP)
  • A petition to boycott the International Congress of Mathematicians in Philadelphia, citing US visa barriers and policies, has grown to 2,300 signatories from 80 countries โ€” and organizers say Chinese mathematicians could be the decisive bloc. (Source: SCMP)

What to Watch

  • DeepSeek V4 enterprise adoption. With vision capabilities in beta and price cuts already rolled out, watch for early production deployments outside China and independent benchmark verification of V4’s coding and agentic claims.
  • China AI chip market acceleration. The Cambricon-SMIC-Hua Hong pipeline is now directly tied to DeepSeek’s release cadence. Supply chain watchers should track whether V4-driven demand shortens Chinese fab timelines.
  • Sovereign AI model exports. HKGAI’s overseas expansion plans for its DeepSeek-based, Huawei-chip-optimized stack are the first test of whether Chinese AI models can gain traction in non-China markets as a sovereign AI proposition.
  • Court ruling ripple effects. The Hangzhou precedent on AI-driven termination could influence how Chinese enterprises approach workforce automation โ€” and may signal broader Chinese labor protections that multinationals need to factor into automation plans.

That is this week’s China AI Weekly โ€” a V4-shaped week where one model launch simultaneously advanced China’s AI capabilities, reshaped chip market expectations, and drew direct top-level policy attention. The question for enterprise teams evaluating these models is no longer “can they compete?” โ€” it is “how fast can the ecosystem around them scale?”

Navigating the China AI landscape for your enterprise architecture? Big Hat Group delivers enterprise AI consulting and Azure-native deployments built on vendor-neutral orchestration and documented governance. Whether you are evaluating DeepSeek V4, assessing chip supply chain risk, or building a multi-model fallback strategy, book a discovery call to scope the work. It takes 30 minutes to get a qualified opinion โ€” not a sales pitch.

Sources:

  • SCMP: “Who could gain from DeepSeek’s V4 with China chips poised for stronger demand?” (2026-05-04)
  • SCMP: “DeepSeek adds AI vision in major move: ’the whale can now see’” (2026-04-29)
  • SCMP: “Hong Kong puts its own spin on DeepSeek with China-chip AI push abroad” (2026-05-04)
  • SCMP: “Xi Jinping urges ‘disruptive innovation’ to boost China amid high-stakes US tech race” (2026-04-30)
  • SCMP: “AI cost-cutting not a legal excuse to fire workers, Chinese court says” (2026-05-04)
  • SCMP: “AI helped Shenzhen judges handle cases 50% faster. Is this the future for China?” (2026-05-04)
  • SCMP: “Oxford AI star Song Yuhang returns to China, but why did he leave chip start-up?” (2026-04-30)
  • SCMP: “Could Chinese mathematicians tip the balance in a US conference boycott?” (2026-05-01)