DeepSeek makes a comeback with V4-Pro and V4-Flash, one year following its ‘Sputnik moment’.

      The startup from Hangzhou introduced preview versions of both models on Hugging Face on Friday. The V4-Pro model is said to excel in coding and mathematics among open models, ranking just behind Gemini 3.1-Pro in terms of world knowledge, and is considered to be “marginally short” of GPT-5.4 and Gemini 3.1-Pro, a disparity DeepSeek estimates to be “approximately 3 to 6 months.” Both models are open-source.

      DeepSeek, the AI startup based in Hangzhou that shook the Silicon Valley scene with its R1 model in January 2025, unveiled preview versions of its new flagship models on Friday, nearly a year after the initial upheaval.

      The company presented DeepSeek-V4-Pro and DeepSeek-V4-Flash on Hugging Face, labeling the launch as the most powerful open-source AI platform available and a direct competitor to companies like OpenAI and Anthropic.

      Both models adhere to the open-source framework established by their predecessors, allowing developers the freedom to use and modify the source code.

      A significant technical advancement in V4 is the Hybrid Attention Architecture, which DeepSeek claims enhances the model’s ability to maintain context during extended conversations. This architecture is paired with a one-million-token context window, capable of processing entire codebases or lengthy documents in a single prompt, and is intended for tasks requiring agency and long-term reasoning, where earlier models experienced diminished performance as context length increased. The Flash variant is optimized for speed and cost-effectiveness, while the Pro variant focuses on maximizing capability.

      DeepSeek’s own evaluations categorize V4-Pro as the leading open-source model in coding and mathematics, with performance only trailing Google’s proprietary Gemini 3.1-Pro for world knowledge.

      When compared to the current benchmarks for closed-source models, including OpenAI’s GPT-5.4 and Gemini 3.1-Pro, the company asserts that V4-Pro falls “marginally short” and openly assesses that its “developmental trajectory” is behind state-of-the-art frontier models by about 3 to 6 months.

      This approach is atypical in AI model releases, which usually highlight areas where the new model excels. DeepSeek’s decision to provide a gap estimation rather than asserting equivalence may indicate a rare level of transparency or a strategic effort to set realistic expectations prior to independent evaluations.

      The narrative surrounding chips introduces a politically sensitive aspect. DeepSeek collaborated with Chinese AI chip manufacturers Huawei and Cambricon to tailor V4 for their latest hardware, according to Reuters, citing The Information.

      The company did not provide early access for optimization to Nvidia or AMD, diverging from the usual industry practice where Western chipmakers typically receive new model weights first for refining performance. Running a cutting-edge model of this scale on Huawei’s Ascend chips, instead of Nvidia H100s or H200s, would serve as significant evidence for China's domestic AI hardware supply chain, which has faced US export restrictions since October 2022. While the V4 launch does not eliminate these geopolitical challenges, it tests them in a commercially significant way.

      The choice of release timing, on Friday, April 24, 2026, roughly one year after DeepSeek-R1, is intentional. The January 2025 launch of DeepSeek-R1 caused Nvidia’s market capitalization to drop by nearly $600 billion in a single day as investors reassessed the computational requirements for developing frontier AI.

      Marc Andreessen referred to it as “AI’s Sputnik moment.” The underlying assertion was that a Chinese lab had matched OpenAI’s leading reasoning model while investing less than $6 million in computing, a figure some analysts contested but which nonetheless reset global expectations regarding the costs of frontier AI development.

      V4 is entering a different market landscape: one where OpenAI has launched GPT-5.5 on the same day, where Anthropic holds a $1 trillion valuation in secondary markets, and where the US-China AI rivalry has become a clear factor in trade and technology policy. DeepSeek’s second chapter is unfolding in a significantly more competitive environment than its initial foray.

      Both models are preview releases rather than final versions. Independent benchmarking has not yet been finalized, and DeepSeek’s own evaluations should be viewed as preliminary until corroborated by external assessments.

      The same caution applied to R1, whose claims were largely confirmed by outside testing shortly after its release. Whether V4 will withstand similar scrutiny will become apparent within the week.