An estimated 80% of US startups now use Chinese AI base models to develop derivative products for their businesses, according to Andreessen Horowitz cited in a recent US-China Economic and Security Review Commission report. Airbnb uses Alibaba’s Qwen for customer service chatbots. Derivatives of the Qwen model family have surpassed 100,000 on Hugging Face, making it the largest open-weight model ecosystem on the platform — bigger than any Western counterpart, including Meta’s Llama.
If that’s the picture in the US, the adoption story in Southeast Asia is moving even faster — and with fewer of the national security concerns that complicate the conversation in Washington. In a region where cost sensitivity, multilingual requirements, and infrastructure constraints shape every technology decision, Chinese open-source AI models aren’t an ideological choice. They’re a practical one.
What’s actually being deployed, and where?
The concrete examples are more specific than the headlines suggest.
In Singapore, banking giant OCBC has rolled out more than 30 internal tools powered by open-source AI. Staff use Google’s Gemma for document summaries, Alibaba’s Qwen to assist with coding, and DeepSeek to analyse market trends. OCBC’s operations span Singapore, Hong Kong, Malaysia, Indonesia, Thailand, and Vietnam — meaning these tools are running across regulatory environments simultaneously.
In Indonesia, AI venture AIonOS partnered with Indosat Ooredoo Hutchison — one of the country’s major telecoms — to develop solutions powered by DeepSeek. The collaboration targets tourism, the knowledge economy, and sustainable agriculture. This isn’t a pilot. It’s a telco-backed deployment using a Chinese model to address core sectors of the Indonesian economy.
Malaysia’s Communications Ministry launched a sovereign, full-stack AI ecosystem powered by Huawei GPUs. That’s a government choosing Chinese hardware as the foundation for national AI infrastructure — a decision with implications that extend well beyond the technical.
ByteDance’s Doubao chatbot is the most popular AI app in China, with 155.2 million weekly active users (QuestMobile). The company’s $8.8 billion commitment to regional data centre development, with significant investment in Thailand, positions TikTok’s parent company as an AI infrastructure provider across Southeast Asia — not just a content platform.
Why are Chinese models winning on adoption?
Three structural factors explain the pattern.
Cost. Chinese AI models run at roughly one-sixth to one-quarter the cost of comparable American systems, according to a RAND report published in early 2026. DeepSeek trained its V3 model for approximately $6 million, compared to an estimated $100 million for OpenAI’s GPT-4. For enterprises and startups in markets where capital is scarce and margins are thin, that cost differential isn’t marginal — it’s decisive.
Open weights. The Chinese model ecosystem has gone overwhelmingly open-source since DeepSeek’s breakout. Baidu CEO Robin Li had insisted closed-source models would dominate. DeepSeek’s success reversed that position within weeks. Today, Hugging Face is packed with releases from Baidu, ByteDance, Tencent, and startups like Moonshot AI. Open weights mean developers can fine-tune models for local languages, regulatory requirements, and specific use cases without dependency on API access or pricing decisions made in San Francisco.
Multilingual capability. Southeast Asia’s linguistic diversity is extreme. Models trained primarily on English text don’t handle Thai, Vietnamese, Bahasa Indonesia, or Tagalog with the same fluency. Chinese model developers — accustomed to building for a linguistically complex domestic market — have prioritised multilingual training data and cross-language performance in ways that map well to Southeast Asian deployment.
Which Chinese models matter most in the region?
The landscape has moved well beyond a single-company story. Four model families are shaping the regional picture.
DeepSeek remains the most recognisable name. Its R1 model — the one that triggered a $593 billion single-day Nvidia market cap loss — demonstrated that frontier-level AI performance doesn’t require frontier-level spending. DeepSeek’s V4 model launched in February 2026, reportedly optimised for coding workloads and testing competitively against Anthropic’s Claude Sonnet and OpenAI’s GPT-4o on coding benchmarks. The company’s structure matters: it’s funded by a quantitative hedge fund, not public shareholders, which lets it prioritise research over monetisation. DeepSeek services have been heavily adopted in Southeast Asia and Africa.
Alibaba’s Qwen is arguably more significant for enterprise deployment even if it draws fewer headlines. Qwen claims more than 100 million monthly active users. Alibaba updated its Qwen AI app to enable shopping, food ordering, and payments without leaving the app — integrating with Taobao and its broader e-commerce ecosystem. The Qwen 3.5 series is imminent, with improved maths reasoning and coding performance. For Southeast Asian businesses already operating within Alibaba’s commercial ecosystem — and many are — Qwen offers a natural AI layer.
Baidu’s Ernie (including Ernie 4.5 and the X1 reasoning model) is most relevant for companies operating in or targeting the Chinese market. Baidu has integrated DeepSeek’s R1 model into its own search engine and launched Qingduo, an AI-powered creative platform that went from generating roughly 20 ad creatives per hour to more than 2,000 after integrating DeepSeek. For advertisers and brands entering China, Baidu’s AI marketing stack — combining Wenxin, DeepSeek, and the country’s dominant search platform processing over 6 billion daily searches — is the primary interface.
ByteDance operates through multiple AI surfaces: the Doubao chatbot in China, its Seedance video generation models (praised by Chinese state media), and increasingly through TikTok’s advertising and commerce infrastructure in Southeast Asia. ByteDance’s AI play in the region is inseparable from its data centre investments and its existing distribution through TikTok and TikTok Shop.
Beyond these four, Zhipu AI’s GLM models and Moonshot AI’s Kimi series are gaining traction. Moonshot’s Kimi K2.5, revealed in January 2026, claimed video-generation and agentic capabilities outperforming the three leading US models.
What are the security and governance implications?
Adoption is outrunning governance. The US National Institute of Standards and Technology found that DeepSeek’s open models presented certain security vulnerabilities. US policymakers have raised concerns about data flows, model behaviour, and the strategic implications of building American businesses on Chinese AI foundations.
In Southeast Asia, these concerns are present but weighted differently. For most enterprises in the region, the immediate question isn’t geopolitical — it’s whether the model works, how much it costs, and whether it handles local languages adequately. Governments like Malaysia’s are actively choosing Chinese hardware infrastructure for sovereign AI programmes. Singapore’s OCBC — a regulated financial institution — is deploying Chinese models alongside Western ones in the same internal toolkit.
The more consequential governance question may be about dependency. If the open-weight model ecosystem becomes predominantly Chinese in origin — and the Hugging Face data suggests it’s heading that direction — then the building blocks for AI applications globally, including in Southeast Asia, increasingly originate in China. That’s not a security alarm in the same way as, say, a compromised supply chain. But it does mean that the foundational AI layer for a growing share of the global economy is shaped by Chinese research priorities, training methodologies, and — potentially — values embedded in training data.
What does this mean for businesses operating in Southeast Asia?
The practical takeaway is straightforward: Chinese AI models are already a significant part of the regional technology stack, and their share is growing. Companies operating in Southeast Asia should be evaluating these models alongside Western alternatives — not out of ideological preference, but because ignoring them means missing cost advantages, multilingual capabilities, and integration pathways that matter in this market.
The landscape is evolving fast. DeepSeek, Alibaba, ByteDance, and their competitors are releasing updated models quarterly. The right approach isn’t to pick one model family and commit — it’s to build evaluation and deployment pipelines that can accommodate models from multiple providers, adjust as capabilities change, and maintain flexibility as regulatory requirements develop across the region’s diverse markets.
The Chinese AI model story in Southeast Asia isn’t about the future. It’s about what’s already running in production systems across Singapore, Indonesia, Malaysia, and Thailand today.
