DeepSeek on Friday released preview versions of its new flagship open source AI model family, V4, announcing two distinct configurations tailored for different scale and efficiency needs.
According to a post on the Hugging Face platform, the company has made available DeepSeek-V4-Pro and DeepSeek-V4-Flash. The Pro variant carries 1.6 trillion parameters, while the Flash edition is a smaller, more compact model with 284 billion parameters. DeepSeek positioned the series as providing enhanced reasoning and overall performance compared with prior open source builds.
DeepSeek said that its most capable offering, V4-Pro-Max, attains what the firm described as "top-tier performance in coding benchmarks" and notably narrows performance gaps with prominent closed-source models on tasks that test reasoning and agentic capabilities. The company provided comparative results on the MMLU-Pro coding benchmark showing DeepSeek V4-Pro matching OpenAI's GPT-5.4 and slightly trailing two other leading closed-source models - Google’s Gemini-3.1-Pro and Anthropic’s Claude Opus 4.6 - based on the data DeepSeek published.
Friday's V4 preview is the company's first substantial, ground-up model launch since the R1 release in early-2025. R1 had been seen as a significant advance in open source AI, delivering performance that many thought comparable to closed-source offerings. That earlier rollout initially coincided with deep declines in global technology stocks, as investors reassessed demand for large new AI infrastructure when more efficient models could offer similar output.
The V4 announcement arrives shortly after media reports indicated that Chinese technology groups Tencent and Alibaba were in discussions to invest in DeepSeek at a valuation in excess of $20 billion. DeepSeek is counted among a group of six high-profile Chinese AI unicorns sometimes referred to as the country's "AI Tigers," companies viewed as central to national ambitions in the fast-growing AI sector.
The company framed V4 as an advancement in both scale and efficiency, offering a spectrum from the large-scale Pro model to the leaner Flash edition. DeepSeek's published benchmark comparisons and the timing of reported investor interest underscore the commercial and market attention surrounding the new release, though the company released the preview builds for wider technical evaluation rather than as final commercial products.
Summary - DeepSeek released preview versions of V4, including a 1.6 trillion-parameter Pro model and a 284 billion-parameter Flash model, and highlighted competitive benchmark results for V4-Pro-Max against major closed-source rivals. This is the firm's first major ground-up model release since R1 in early-2025 and follows reports of potential investment talks with Tencent and Alibaba valuing the company at over $20 billion.