Skip to main content
Live
Main content
Guide · Evergreen

Asian AI

How China, South Korea, Japan, and Southeast Asia are building AI — the labs, the apps, and the regulatory posture that defines the region in 2026.

Launched · Updated

Asian AI is the half of the global frontier most US-published AI coverage doesn't touch. The Chinese ecosystem alone — Baidu, ByteDance, Tencent, Alibaba, DeepSeek, Moonshot, Zhipu — ships more models, at lower price points, with broader consumer adoption inside China than the US frontier ships in its home market. South Korea, Japan, and Southeast Asia each run their own AI economies with distinct regulatory frames, distinct language priorities, and distinct competitive dynamics that don't show up in the OpenAI-vs-Anthropic narrative US journalism keeps reaching for.

This page is the living reference to Asian AI on AI Chat Daily — the labs, the consumer products, the regulatory shape, and every story we publish tagged Asia AI. We update it whenever a meaningful release lands.

As of April 2026, the most-watched stories in the region are: DeepSeek's continued lead on cost-efficient frontier reasoning, Yuanbao's hybrid Hunyuan+DeepSeek routing strategy, Quark overtaking Baidu in Chinese AI search, and Korea's AI Basic Act setting the most comprehensive regulatory frame in Asia.

Timeline of major releases

  1. Mar 16, 2023
    Baidu launches ERNIE Bot in China
    Baidu becomes the first Chinese tech major to ship a public ChatGPT competitor. Initial response is mixed; Baidu iterates aggressively over the following 18 months.
  2. Aug 15, 2023
    China's Generative AI Interim Measures take effect
    The Cyberspace Administration of China issues registration requirements for any consumer-facing generative AI service. Foundation models must align with "core socialist values" — a regulatory posture that shapes every Chinese model release that follows.
  3. May 15, 2024
    ByteDance's Doubao launches publicly
    TikTok parent ByteDance ships Doubao at aggressive pricing, undercutting OpenAI by 99% in token costs. Triggers China's "AI price war" of summer 2024.
  4. Dec 26, 2024
    DeepSeek-V3 publishes open weights
    The Hangzhou hedge-fund-affiliated lab releases a 671B-param mixture-of-experts model trained for ~$6M, with reasoning quality comparable to GPT-4. Becomes the catalyst for cost-efficient frontier training globally.
  5. Jan 20, 2025
    DeepSeek-R1 reasoning model launches
    Free, open-weight reasoning model that matches OpenAI o1 on math and coding benchmarks. The "DeepSeek moment" reframes the trillion-dollar AI capex debate.
  6. Sep 4, 2025
    Quark passes 100M monthly active users
    Alibaba's AI search assistant — built on Qwen — becomes the highest-MAU consumer AI app in mainland China, ahead of Baidu and Tencent's Yuanbao.
  7. Feb 22, 2026
    Tencent unifies Hunyuan + DeepSeek inside Yuanbao
    Yuanbao adopts a hybrid routing layer that selects between Tencent's Hunyuan models and DeepSeek for different query types. First major Chinese consumer assistant to ship multi-lab routing as a default.
  8. Apr 15, 2026
    South Korea's AI Basic Act enters force
    Korea passes its comprehensive AI law modeled partly on the EU AI Act but with carve-outs for the local chaebols. Naver and Kakao adapt their consumer AI products to the new compliance regime.

The Chinese frontier in 2026

The Chinese AI frontier is shaped by four constraints that the US frontier doesn't share: export controls on the highest-end NVIDIA chips, mandatory model registration with the Cyberspace Administration of China, the requirement that consumer outputs align with state values, and a domestic talent pool that has stayed home (relatively few of the top Chinese AI researchers have decamped to the US since 2022).

Out of those constraints have come the products that increasingly define the global open-weights landscape. DeepSeek's V3 and R1 are the cost-efficient reasoning models that the rest of the industry now benchmarks against. Alibaba's Qwen has become the most-downloaded open-weights model family on Hugging Face, ahead of Llama on most weeks. Tencent's Hunyuan, ByteDance's Doubao, and Moonshot's Kimi each have meaningful consumer adoption inside China — Doubao reportedly has the largest paid-API revenue of any Chinese consumer AI as of Q1 2026.

For Western readers, the practical question is when (and which) Chinese open-weights models to use. The answer in 2026 is "DeepSeek for reasoning, Qwen for everything else." Both ship under permissive licenses, both run on Western inference providers, and both produce frontier-grade output at fractions of the OpenAI/Anthropic per-token cost.

Korean AI: Naver, Kakao, Samsung

South Korea's AI ecosystem is structured around the chaebols. Naver runs HyperCLOVA, the Korean-language frontier that powers the AI features inside Naver search and shopping. Kakao runs Kanana, which ships inside KakaoTalk — the dominant Korean messaging app — as a Korean-language reasoning assistant. Samsung integrates its Gauss models into Galaxy phones via Galaxy AI, which runs on-device for many tasks and routes to cloud Gemini for harder ones.

What makes Korean AI distinct is the regulatory frame. Korea's AI Basic Act passed in late 2024 and took effect in April 2026. It's the most comprehensive AI law in Asia after China's, with explicit obligations around high-risk uses, mandatory pre-deployment evaluations, and data-protection provisions that go beyond Korea's existing PIPA framework. The chaebols have lobbied for and received some carve-outs, but the regime is meaningfully stricter than anything Japan or Southeast Asia has shipped.

For most foreign readers, the relevance of Korean AI is twofold: HyperCLOVA and Kanana are the strongest Korean-language LLMs in the world (Western models still trail on Korean reasoning), and Samsung's on-device AI work foreshadows where Apple Intelligence is headed in 2026-2027.

Japan and Southeast Asia: smaller scale, distinct strategy

Japan's frontier is led by Sakana AI, the Tokyo-based research lab founded by ex-Google researchers David Ha and Llion Jones. Sakana's contribution to global AI is research-heavy: evolutionary model merging, novel optimization approaches, and the most-cited Japanese AI papers of 2024-2026. NTT and Fujitsu ship enterprise LLMs targeted at Japanese-language workloads. Rakuten's AI-7B family is open-weight and Japanese-tuned.

Southeast Asia's most important contribution is the SEA-LION model family from AI Singapore, which targets Southeast Asian languages (Bahasa Indonesia, Thai, Vietnamese, Tagalog, Filipino, Malay) better than any Western model. Indonesia's GovTech ships its own civic-services LLM. Vietnam's VinAI does the same for Vietnamese. Singapore is the regulatory hub for the major US labs in the region — every frontier US lab has its APAC compliance and sales operation in Singapore.

For readers building products that need to work in Asian-language consumer markets, the practical advice in 2026 is to test SEA-LION for Southeast Asian use cases, Qwen for Chinese, HyperCLOVA or Kanana for Korean, and Sakana's models or Rakuten AI-7B for Japanese — and to benchmark them honestly against Claude/GPT/Gemini on the specific language tasks you care about. The gap on minority languages is often larger than the gap on English benchmarks suggests.

How we cover Asian AI

Asian AI is a deliberate beat for AI Chat Daily, not an afterthought. Our editorial calendar reserves 2-4 tool review slots per week for Asian, Korean, or Japanese AI products. Every Asian AI release of meaningful scale gets covered the same week it ships. We track the Chinese AI price wars, the Korean regulatory frame, and the on-device Galaxy AI roadmap as ongoing stories rather than one-off reports.

The audience for this beat is largely Western: developers, founders, researchers, and journalists trying to understand a frontier their normal information diet underweights. The pages on this site that pull the strongest CTR from Singapore, Malaysia, and Hong Kong are the Asian AI tool reviews — confirming that the appetite is there.

Our coverage

Everything tagged Asia AI appears in Latest news above. For specific products, see Doubao, Yuanbao, Quark, Cici, QQ Browser, Baidu AI Search, Hi Translate, Papago, Meituan, and ADot. For the open-weights Chinese frontier, see the DeepSeek hub.

Frequently asked questions

Which Chinese AI model is the strongest in 2026?
DeepSeek-R1 and DeepSeek-V3 lead on most reasoning benchmarks among Chinese frontier models, with Qwen 3 and Hunyuan close behind. ByteDance's Doubao is the fastest-growing by usage but trails on pure capability. The cost-efficiency story tilts strongly toward DeepSeek across both API and open-weights deployments.
How does Chinese AI regulation differ from the US and EU?
China's Generative AI Interim Measures (effective August 2023) require pre-launch registration with the Cyberspace Administration of China, content alignment with state values, watermarking on AI-generated content, and explicit user consent for training. Compared to the EU AI Act, the Chinese regime is more prescriptive about content but lighter on use-case risk classification. The US has no analogous federal regime.
What is Yuanbao?
Yuanbao is Tencent's consumer AI assistant, launched in 2024. It originally ran on Tencent's Hunyuan models and added DeepSeek as a hybrid backend in 2025. Yuanbao is one of the top three consumer AI apps in mainland China alongside Doubao and Quark.
What is Doubao?
Doubao is ByteDance's consumer AI app and API, launched May 2024 at aggressive pricing that undercut OpenAI by ~99% per token. Its public release triggered China's domestic AI price war. ByteDance also offers Doubao via API for developers and enterprises.
What is Quark?
Quark is Alibaba's AI search assistant. Built on the Qwen model family, it crossed 100M monthly active users in late 2025 and overtook Baidu as the most-used consumer AI search product in mainland China. Quark is bundled with Alibaba's mobile ecosystem.
Are Chinese AI models available outside China?
Open-weights models like DeepSeek-V3, DeepSeek-R1, and Qwen 3 are downloadable from Hugging Face and run on Western inference providers (Together, Hyperbolic, Fireworks). Hosted Chinese AI products (Doubao, Yuanbao, Quark) are not officially available outside mainland China, though developers can call ByteDance's Doubao API from international IPs in some cases.
What's happening with AI in South Korea?
Naver's HyperCLOVA family powers the search-and-shopping AI features inside Naver. Kakao's Kanana model ships inside KakaoTalk and was the first Korean LLM with substantive Korean-language reasoning ability. Samsung is integrating its Gauss models across Galaxy phones. Korea's AI Basic Act (effective 2026) is the most comprehensive regulatory regime in Asia outside China.
What's happening with AI in Japan?
Sakana AI (Tokyo) leads the research side, with novel approaches to evolutionary model merging. NTT, Fujitsu, and Rakuten each ship their own LLMs targeted at Japanese-language enterprise use cases. Japan's regulatory posture is the loosest in the region — no AI-specific law as of 2026, only guidance from METI.
What about Southeast Asia?
Singapore's AI Singapore initiative funds the SEA-LION model family, which targets Southeast Asian languages (Bahasa Indonesia, Thai, Vietnamese, Tagalog) better than Western models. SEA-LION is open-weight. Indonesia's GovTech runs its own LLM for civic services. The major US labs all treat Singapore as their APAC hub for compliance + sales.
How do Chinese AI prices compare to US prices?
Chinese frontier APIs are generally 80-99% cheaper per million tokens than the US frontier. As of April 2026, DeepSeek-V3 is roughly $0.27 input / $1.10 output per million tokens; ByteDance Doubao is similar. The price gap drives most of the global open-weights deployment volume to Chinese-trained models.

Related coverage

Go deeper
    About this guide
    This is an evergreen reference page maintained by the AI Chat Daily newsroom. We update it whenever Asian AI ships a new release, changes pricing, or generates news that affects how you use it.
    AI Box Daily briefingFree · Daily · No fluff

    Stay ahead of everyone in AI.

    The tightly edited AI news email engineers, founders, and investors actually open. One email. Every weekday. Five minutes to finish.

    Loved by 10,000+ AI professionals
    Free forever. Unsubscribe with one click.

    The briefing read inside teams at