Source Credit : Portfolio Prints
DeepSeek on Friday unveiled a preview of its highly anticipated V4 large language model, opening it up for early testing and signaling its next push into the rapidly evolving global AI race.
The release comes more than a year after the Hangzhou-based startup shook markets with its R1 reasoning model, which stunned industry observers with its performance despite dramatically lower development costs.
In keeping with its strategy, DeepSeek has made V4 open-source, allowing developers to download, run, and modify the model locally—an approach that continues to differentiate it from more closed ecosystems built by Western rivals.
The model is offered in both “Pro” and “Flash” variants, tailored to different compute needs. DeepSeek claims V4 delivers strong performance against domestic peers, particularly in agent-based workflows, knowledge processing, and inference tasks—areas increasingly central to real-world AI deployment.
“DeepSeek’s V4 preview is a serious flex,” said Neil Shah of Counterpoint Research, pointing to notably lower inference costs compared to earlier models. Inference costs—the expense of running trained models in production—are emerging as a critical battleground as AI adoption scales.
The company added that V4 has been optimized for compatibility with widely used agent frameworks such as Claude Code and OpenClaw, reinforcing its positioning in the growing “agentic AI” ecosystem.
According to Counterpoint analyst Wei Sun, benchmark indicators suggest V4 could deliver excellent agent capability at significantly lower cost, underscoring its potential appeal to developers prioritizing efficiency.
Founded in 2023, DeepSeek rose to prominence with its open-source V3 model in late 2024, followed shortly by the release of R1 in January 2025—a reasoning model that matched or exceeded leading systems from OpenAI and Google across several benchmarks.
R1, in particular, rattled investors after the company disclosed it had been built in just two months for under $6 million using lower-capacity Nvidia chips—raising questions about the efficiency of Big Tech’s multibillion-dollar AI spending and the durability of U.S. leadership.
Since then, DeepSeek has rolled out incremental updates, though none have matched the disruptive impact of R1. Analysts expect V4 to be less market-shaking, as investors have already adjusted to the reality that Chinese AI models can compete on both performance and cost.
“Traders have already priced this in,” said Ivan Su of Morningstar, noting that V4 reframes competition by positioning domestic open-source models directly against each other—a sign of intensifying rivalry within China.
Competition in China’s AI sector has accelerated sharply, with major players such as Alibaba and ByteDance launching new models this year.
Market reaction reflected that pressure: shares of AI firms including MiniMax and Zhipu fell about 8% in Hong Kong trading, while Hangzhou-based Manycore Tech dropped roughly 9%.
A central question surrounding V4 is its underlying hardware. Huawei confirmed that its latest AI computing cluster, powered by Ascend processors, can support the model, though the extent of domestic chip usage in training remains unclear.
Chinese developers continue to face restrictions on acquiring Nvidia’s most advanced chips due to evolving U.S. export controls, prompting Beijing to accelerate efforts to build a self-reliant semiconductor ecosystem.
Wei Sun noted that V4’s ability to run natively on local chips could have far-reaching implications, strengthening China’s push for AI sovereignty while reducing dependence on foreign hardware.
Investor sentiment toward China’s chip sector, however, was more upbeat. Following the announcement, shares of contract manufacturers SMIC and Hua Hong Semiconductor surged 9% and 15%, respectively, reflecting optimism about rising domestic demand.
