Chinese AI models surge as DeepSeek prepares V4 launch amid open-source dominance

What's next for Chinese open-source AI

Just last week the Chinese firm Moonshot AI released its latest open-weight model, Kimi K2.5, which came close to top proprietary systems such as Anthropic's Claude Opus on some early benchmarks. The difference: K2.5 is roughly one-seventh Opus's price. On Hugging Face, Alibaba's Qwen family -- after ranking as the most downloaded model series in 2025 and 2026 -- has overtaken Meta's Llama models in cumulative downloads. And a recent MIT study found that Chinese open-source models have surpassed US models in total downloads. For developers and builders worldwide, access to near-frontier AI capabilities has never been this broad or this affordable. These models differ in a crucial way from most US models like ChatGPT or Claude, which you pay to access and can't inspect. The Chinese companies publish their models' weights -- numerical values that get set when a model is trained -- so anyone can download, run, study, and modify them. If open-source AI models keep getting better, they will not just offer the cheapest options for people who want access to frontier AI capabilities; they will change where innovation happens and who sets the standards. Here's what may come next. When DeepSeek launched R1, much of the initial shock centered on its origin. Suddenly, a Chinese team had released a reasoning model that could stand alongside the best systems from US labs. But the long tail of DeepSeek's impact had less to do with nationality than with distribution. R1 was released as an open-weight model under a permissive MIT license, allowing anyone to download, inspect, and deploy it. On top of that, DeepSeek also published a paper detailing its training process and techniques. For developers who access models via an API, DeepSeek also undercut competitors on price, offering access at a fraction the cost of OpenAI's o1, the leading proprietary reasoning model at the time. Within days of its release, DeepSeek replaced ChatGPT as the most downloaded free app in the US App Store. The moment spilled beyond developer circles into financial markets, triggering a sharp sell-off in US tech stocks that briefly erased roughly $1 trillion in market value. Almost overnight, DeepSeek went from a little-known spin-off team backed by a quantitative hedge fund to the most visible symbol of China's push for open-source AI. China's decision to lean into open source isn't surprising. It has the world's second-largest concentration of AI talent after the US. plus a vast, well-resourced tech industry. After ChatGPT broke into the mainstream, China's AI sector went through a reckoning -- and emerged determined to catch up. Pursuing an open-source strategy was seen as the fastest way to close the gap by rallying developers, spreading adoption, and setting standards.

A year on from DeepSeek shock, get set for flurry of low-cost Chinese AI models

BEIJING, Feb 12 (Reuters) - One year after Chinese startup DeepSeek rattled the global tech industry with the release of a low-cost artificial intelligence model, its domestic rivals are better prepared, vying with it to launch new models, some designed with more consumer appeal. The Hangzhou-based firm's meteoric rise in early 2025, during China's Spring Festival holiday, upended China's AI industry, pushing low-cost, open-source models to the forefront of its AI ecosystem. This time, DeepSeek will be joined by several other firms in launching new products around China's longest and busiest holiday period, which officially begins on February 15. While the industry was stunned when DeepSeek broke through with a strong AI model despite U.S. export controls restricting access to advanced semiconductors, now the market wants to see what Chinese companies come up with next, said Alfredo Montufar-Helu, a managing director at Ankura Consulting in Beijing. "The surprise would be if some of these new models end up being underwhelming. I think there are high expectations here," he said. Zhipu AI (2513.HK), opens new tab on Wednesday released its latest AI model, which it said features enhanced coding capabilities and the ability to perform long-running tasks without any user prompts. ByteDance on Thursday officially unveiled Seedance 2.0, a video generation AI model "capable of producing cinematic blockbusters in seconds", according to the Chinese state-backed Global Times newspaper. ByteDance is also expected to roll out upgrades to its Doubao chatbot, currently China's most popular AI app with 155.2 million weekly active users, according to QuestMobile. DeepSeek, too, is preparing to release its next-generation model V4 and rival Alibaba (9988.HK), opens new tab is expected to unveil its Qwen 3.5 series, featuring improved mathematical reasoning and coding capabilities, tech industry news site The Information reported last month. Qwen developers submitted support code earlier this month for "upcoming Qwen 3.5 series models" to the open-source repository Hugging Face, typically a sign that a release is imminent. Alibaba, ByteDance and DeepSeek have not announced a formal release date for their upgraded models. The companies did not respond to a request for comment. LOW-COST, OPEN SOURCE NOW THE NORM DeepSeek's initial release in January 2025 triggered a global tech selloff and wiped $593 billion from AI chipmaker Nvidia's (NVDA.O), opens new tab market value in a single day and spurred its Chinese rivals to release upgrades to their own models. In the past two years, DeepSeek's models have repeatedly undercut competitors' prices, pushing usage costs significantly below many U.S. offerings. In the U.S., investors saw DeepSeek's claim it had built a model comparable to OpenAI's best but at a fraction of the cost, as a challenge to the assumption that only companies spending tens of billions of dollars on computing infrastructure could produce cutting-edge AI. A report by research group RAND on U.S.-China AI competition published last month found that Chinese models operate at roughly one-sixth to one-fourth the cost of comparable U.S. systems. "DeepSeek showed the industry that you can create a very good model even when you're resource-constrained," said Lian Jye Su, chief analyst at tech research firm Omdia. "The combination of open-source access, strong reasoning capabilities and low deployment costs has become a defining model for how Chinese vendors now approach foundation models." Before DeepSeek's breakout, some Chinese industry leaders, including Baidu CEO Robin Li, had argued that closed-source systems would dominate. Within days of DeepSeek's assistant overtaking ChatGPT in Apple's App Store downloads in the U.S., Baidu and other leading firms began opening portions of their own models. Hugging Face is now dominated by releases from Chinese tech giants such as Baidu (9888.HK), opens new tab, ByteDance and Tencent (0700.HK), opens new tab, and startups such as Moonshot. "Chinese companies are actively embracing open source, significantly lowering the barriers for global developers and enterprises to access cutting-edge AI technology," Global Times wrote in a Wednesday editorial praising Seedance 2.0. Besides adopting DeepSeek's open-source approach, competitors have also stepped up recruitment of top AI researchers. IMITATION AND DIVERGENCE While DeepSeek remains focused on advancing core model performance, rivals are shifting emphasis toward integrating AI into consumer services. Alibaba's Qwen chatbot has recently experimented with enabling users to purchase goods directly through conversational prompts. The pivot reflects commercial realities. Companies such as Alibaba face shareholder pressure to monetise AI investments through consumer and enterprise applications while continuing to fund expensive infrastructure expansion. DeepSeek remains structurally distinct. Its parent company is a quantitative hedge fund controlled by founder Liang Wenfeng, allowing it to prioritise research over commercialisation and avoid external investor pressure. Reporting by Eduardo Baptista; Editing by Miyoung Kim and Sonali Paul Our Standards: The Thomson Reuters Trust Principles., opens new tab * Suggested Topics: * Artificial Intelligence Eduardo Baptista Thomson Reuters Eduardo Baptista is a Senior Correspondent for Reuters based in Beijing, covering China's technology, space, and automotive industries. He has led enterprise and investigative reporting on China's military-linked companies, artificial intelligence and semiconductor supply chains, as well as macroeconomic and industrial policy. Baptista has reported from China for nearly a decade and holds a BA in History from the University of Cambridge.

DeepSeek V4 release coming soon: What to expect from the Chinese AI star that rattled markets last year

DeepSeek V4 features and upgrades: As China celebrates the Lunar New Year, a new wave of artificial intelligence launches is building momentum. Companies like Alibaba, ByteDance, and Zhipu have unveiled new products in the run-up to the holiday. But much of the attention is centered on DeepSeek, the open-source AI firm whose last major release shook global markets last year. According to The Information, DeepSeek is expected to release its fourth-generation model, V4, around the Lunar New Year. If it launches as anticipated, V4 could replace the V3 model that powered the assistant app which became the top-rated free application in the United States and worldwide, overtaking OpenAI's ChatGPT last January. DeepSeek's V3 was not just another model update. It rattled global markets and triggered a sell-off in US-led tech stocks. Industry giants including Nvidia saw shares plunge 17%, wiping out $600 billion in market value before recovering later in the day. That market reaction cemented DeepSeek's reputation as a serious global contender. Also read: Gold price jumps above $5,000, silver price surges 6% as Fed Minutes and global talks drive market volatility - will gold hit $6,000 and silver reach $309 Even before V4's official release, anticipation has been building. The company recently upgraded its chatbot's context window, expanding how much information it can process and remember in a single task, according to the South China Morning Post. For users, that means longer conversations and more complex tasks handled in one go. However, DeepSeek's rapid rise has not been without controversy. Countries including Italy, Denmark, and Czech Republic have banned government agencies from using its models over data security and cybersecurity concerns. Local media in Belgium reported that officials stopped using DeepSeek in December. DeepSeek's market position has shifted dramatically over the past year. Its share of the open-source model market fell from 50% at the start of 2025 to under 25% by year-end. Intensifying competition from Qwen, Kimi K2, and InternLM has chipped away at its dominance. At the same time, DeepSeek redirected resources toward building a broader "model + tools" ecosystem and preparing its next-generation architecture. That makes V4 more than a routine update. It is a pivotal moment for the company, a bid to reclaim technical leadership and validate its strategic pivot. Also read: Scientists create a living 3D 'heart-on-a-chip' to study and stop deadly heart disease DeepSeek V4 introduces four major technical upgrades: MODEL1 architecture with tiered KV cache storage aims to cut memory use by 40% by intelligently distributing data across GPU, CPU, and disk storage, as per a Vertu report. Instead of compressing information, the system places frequently used data in faster storage while archiving less-used data elsewhere. This could extend context windows beyond 1 million tokens and significantly reduce costs. Sparse FP8 decoding brings a hybrid precision approach. By calculating critical tokens at higher precision and less important ones at lower precision, V4 reportedly achieves a 1.8x inference speedup with minimal accuracy loss. For high-traffic systems, that translates into meaningful cost savings. Engram memory modules separate short-term context from long-term memory. Rather than reprocessing entire conversation histories every time, key information is stored and selectively retrieved when relevant. This design enables more personalized, context-aware AI agents while controlling costs. mHC optimized residual connections allow the model to learn which layers are most important during training. By introducing a scaling parameter, the system improves training efficiency by 30% and slightly boosts performance benchmarks. DeepSeek is also shifting strategy. Instead of remaining purely a model provider, it is building a China-focused alternative to Cursor, the AI coding tool valued at over $2 billion in 2025. The move reflects a broader transition from infrastructure to application-layer tools and ecosystem development. The company is positioning itself to leverage its proprietary models, localization strengths for Chinese developers, and established community presence. But competition is fierce, and execution will be critical. DeepSeek's trajectory mirrors a broader pattern seen in the AI industry: moving from research breakthroughs to full-fledged platforms. With V4, the company is attempting to regain lost market share, respond to rising competition, and prove that its pivot toward tools and ecosystem development can succeed. When is DeepSeek V4 expected to launch? It is expected to be released around the Lunar New Year, according to The Information. What happened to DeepSeek's market share? It fell from 50% at the start of 2025 to under 25% by year-end.

A year on from DeepSeek shock, get set for flurry of low-cost Chinese AI models - The Economic Times

One year after Chinese startup DeepSeek rattled the global tech industry with the release of a low-cost artificial intelligence model, its domestic rivals are better prepared, vying with it to launch new models, some designed with more consumer appeal. The Hangzhou-based firm's meteoric rise in early 2025, during China's Spring Festival holiday, upended China's AI industry, pushing low-cost, open-source models to the forefront of its AI ecosystem. This time, DeepSeek will be joined by several other firms in launching new products around China's longest and busiest holiday period, which officially begins on February 15. While the industry was stunned when DeepSeek broke through with a strong AI model despite U.S. export controls restricting access to advanced semiconductors, now the market wants to see what Chinese companies come up ⁠with next, ⁠said Alfredo Montufar-Helu, a managing director at Ankura Consulting in Beijing. "The surprise would be if some of these new models end up being underwhelming. I think there are high expectations here," he said. Zhipu AI on Wednesday released its latest AI model, which it said features enhanced coding capabilities and the ability to perform long-running tasks without any user prompts. ByteDance on Thursday officially unveiled Seedance 2.0, a video generation AI model "capable of producing cinematic blockbusters in seconds", according to the Chinese state-backed Global Times newspaper. ByteDance is also expected to roll out upgrades to its Doubao chatbot, currently China's most popular AI app with 155.2 million weekly active users, according to QuestMobile. DeepSeek, too, is preparing to release its next-generation model V4 and rival Alibaba is expected to unveil its Qwen 3.5 series, featuring improved mathematical reasoning and coding capabilities, tech industry news site The ⁠Information reported last month. Qwen developers submitted support code earlier this month for "upcoming Qwen 3.5 series models" to the open-source repository Hugging Face, typically a sign that a release is imminent. Alibaba, ByteDance and DeepSeek have not announced a formal release date for their upgraded models. The companies did not respond to ⁠a request for comment. Low cost, open source now the norm DeepSeek's initial release in January 2025 triggered a global tech selloff and wiped $593 billion from AI chipmaker Nvidia's market value in a single day and spurred its Chinese rivals to release upgrades to their own models. In the past two years, DeepSeek's models have repeatedly undercut competitors' prices, pushing usage costs significantly below many US offerings. In the US, investors saw DeepSeek's claim it had built a model comparable to OpenAI's best but at a fraction of the cost, as a challenge to the assumption that only companies spending tens of billions of dollars on computing infrastructure could produce cutting-edge AI. A report by research group RAND on U.S.-China AI competition published last month found that Chinese models operate at roughly one-sixth to one-fourth the cost of comparable U.S. systems. "DeepSeek showed the industry that you can create a very good model even when you're resource-constrained," said Lian Jye Su, chief analyst at tech research firm Omdia. "The combination of open-source access, strong reasoning capabilities and low deployment costs has become a defining model for how Chinese vendors now approach foundation models." Before DeepSeek's breakout, some Chinese industry leaders, including Baidu CEO Robin Li, had argued that closed-source systems would dominate. Within days of DeepSeek's assistant overtaking ⁠ChatGPT in Apple's App Store downloads in the U.S., Baidu and other leading firms began opening portions of their own models. Hugging Face is now dominated by releases from Chinese tech giants such as Baidu, ByteDance and Tencent, and startups such as Moonshot. "Chinese companies are actively embracing open source, significantly lowering the barriers for global developers and enterprises to access cutting-edge AI technology," Global Times wrote in a Wednesday editorial praising Seedance 2.0. Besides adopting DeepSeek's open-source approach, competitors have also stepped up recruitment of top AI researchers. Imitation and divergence While DeepSeek remains focused on advancing core model performance, rivals are shifting emphasis toward integrating AI into consumer services. Alibaba's Qwen chatbot has recently experimented with enabling users to purchase goods directly through conversational prompts. The pivot reflects commercial realities. Companies such as Alibaba face shareholder pressure to monetise AI investments through consumer and enterprise applications while continuing to fund expensive infrastructure expansion. DeepSeek remains structurally distinct. Its parent company is a quantitative hedge fund controlled by founder Liang Wenfeng, allowing it to prioritise research over commercialisation and avoid external investor pressure.