DeepSeek released a preview of its V4 large language model on Friday, more than a year after its R1 reasoning model briefly upended global tech markets. The Hangzhou-based startup made the new model available in two versions — "pro" and "flash" — depending on size, and said it performs strongly against domestic competitors, particularly in agent-based tasks, knowledge processing, and inference.
The release follows DeepSeek's established pattern of open-source distribution. Developers can download the code, run it locally, and modify it, keeping the model accessible outside closed commercial ecosystems. DeepSeek also said V4 has been optimized for use with popular agent tools including Anthropic's Claude Code and OpenClaw.
Early assessments from analysts were positive on capability and cost. "DeepSeek's V4 preview is a serious flex," Neil Shah, vice president of research at Counterpoint Research, told CNBC, pointing to lower inference costs compared to previous models. Inference costs cover the computational and financial expense of running a trained AI model to generate outputs — a key metric for enterprise adoption. Counterpoint's principal AI analyst Wei Sun said V4's benchmark profile suggests "excellent agent capability at significantly lower cost."
The market reaction is unlikely to match the shock of R1's debut in January 2025, when DeepSeek revealed the reasoning model had been built in two months for under $6 million using lower-capacity Nvidia chips. That disclosure rattled investors and raised serious questions about the U.S. advantage in AI infrastructure spending. Since then, traders have adjusted their expectations about Chinese AI competitiveness.
Ivan Su, senior equity analyst at Morningstar, told CNBC that V4 is unlikely to move markets the way R1 did, because investors have already priced in the reality that Chinese AI is both competitive and cheaper to run. But Su noted a meaningful shift in how DeepSeek is positioning the product. Rather than framing V4 against Western rivals, DeepSeek is explicitly targeting other Chinese open-source models as its direct competition. "This is a framing that didn't exist with R1, and that alone tells you how much domestic competition has intensified," Su said.
That competition has grown substantially since R1's release. Alibaba and ByteDance have both launched new models in 2026, and a cluster of smaller Chinese AI developers has emerged as a serious domestic force. Shares of several Chinese AI companies fell in Hong Kong trading Friday following V4's release, with MiniMax and Knowledge Atlas Technology each dropping around 8%.
DeepSeek founded the company in 2023 and first drew wide attention in late 2024 with its V3 model, which it claimed was trained at a fraction of the cost of comparable U.S. models. V4 represents the company's first major architectural release since then, and its focus on agent tasks — where AI models take sequences of autonomous actions — signals where the competitive frontier in AI is currently moving.
