DeepSeek released preview versions of its V4 large language model on Friday, offering developers a chance to test its Pro and Flash variants as competition in China’s AI sector intensifies.
The Hangzhou-based startup claims its V4-Pro model outperforms all rival open-source models in math and coding tasks, trailing only Google’s closed Gemini 3.1-Pro in world knowledge benchmarks.
According to DeepSeek, the V4-Pro’s performance falls only marginally short of OpenAI’s GPT-5.4 and Gemini 3.1-Pro, suggesting it trails state-of-the-art frontier models by three to six months.
The V4-Flash variant provides similar reasoning capabilities to the Pro version but with faster response times and highly cost-effective pricing, the company said.
Both models follow DeepSeek’s open-source approach, allowing developers to download, run, and modify the code locally in most cases.
DeepSeek said V4 has been optimized for use with popular agent tools such as Anthropic’s Claude Code and OpenClaw, enhancing its utility for agent-based tasks.
Counterpoint Research analysts noted the model offers excellent agent capability at significantly lower inference costs compared to previous versions.
Inference costs refer to the computational and financial expenses of running a trained AI model to generate outputs, a key factor in enterprise adoption.
The release comes just over a year after DeepSeek’s R1 reasoning model stunned global tech markets in January 2025 with performance comparable to ChatGPT and Gemini despite reportedly costing less than $6 million to develop.
That earlier claim sparked debate among tech analysts who questioned whether the startup had access to greater funding or more advanced chips than acknowledged.
DeepSeek’s arrival prompted data privacy and censorship concerns in multiple countries, leading to bans or restrictions on its R1 model in US states, Australia, Taiwan, South Korea, Denmark, and Italy.
Morningstar analyst Ivan Su said V4’s debut is unlikely to repeat R1’s market impact because traders have already priced in the reality that Chinese AI is competitive and cheaper to use.
However, Su added that DeepSeek’s latest positioning frames other Chinese open-source models as direct competitors, reflecting intensified domestic rivalry.
Since R1’s release, DeepSeek has faced increased competition from Alibaba and ByteDance, both of which have launched latest AI models this year.
Shares of several other Chinese AI players declined in Hong Kong trading on Friday amid the shifting landscape.
The Stanford AI Index 2026 found that whereas Silicon Valley retains a slight edge in developing the most advanced models, Chinese companies have effectively closed the performance gap with US rivals.
How does DeepSeek-V4 compare to closed models from OpenAI and Google?
According to DeepSeek, the V4-Pro model trails only Google’s Gemini 3.1-Pro in world knowledge among all models and falls marginally short of OpenAI’s GPT-5.4 and Gemini 3.1-Pro in overall performance, suggesting it is three to six months behind the leading frontier models.
Why might V4 not have the same market impact as DeepSeek’s R1 model?
Analysts say V4’s debut is unlikely to shock markets like R1 did because traders have already priced in the expectation that Chinese AI models are competitive and cost-effective to use.
