Back

DeepSeek Previews V4 AI Model Optimized for Huawei Ascend as White House Levels Theft Accusation

DeepSeek released a V4 preview on April 24, describing optimization for Huawei Ascend, while a Polymarket market entered UMA disputes over whether the preview satisfies formal V4 release criteria.

Fact Check
The claim has three main parts, and each is supported by sources gathered in this run. First, DeepSeek releasing a V4 preview on April 24 is supported by Odaily and PANews, both of which explicitly state a DeepSeek-V4 preview went live on April 24 and mention V4-Pro and V4-Flash. Second, optimization or adaptation for Huawei Ascend is supported by PANews, which says the system was validated on Huawei Ascend NPUs, and by Crypto.news, which explicitly says the preview models were optimized for Huawei Ascend. Third, the statement about a Polymarket market entering UMA disputes over whether the preview satisfies formal V4 release criteria is strongly supported by the Polymarket market page itself, which lays out formal resolution criteria, and by Odaily, which reports that the April 24 preview triggered multiple UMA dispute rounds because a preview may not count as an official V4 release under those rules. Reuters corroborates the White House theft-accusation context. Confidence is medium rather than high because the best direct primary source from DeepSeek's own domain was not fetched in this run, and Reuters page fetches returned minimal body text.
    Reference: 12
Summary

DeepSeek released a preview of its V4 model line on April 24, expanding an earlier rollout that included DeepSeek-V4-Pro and DeepSeek-V4-Flash and stating that the model was explicitly optimized for Huawei’s Ascend chip platform. Official materials also said DeepSeek validated its expert parallel scheme on both Nvidia GPUs and Huawei Ascend NPUs, while its API page noted V4-Pro throughput remains limited. The release triggered a Polymarket dispute, with the platform’s DeepSeek V4 market entering three UMA dispute rulings because opponents argue V4-Preview does not meet the market’s formal criteria for a DeepSeek V4 successor and should therefore resolve No. The release also coincided with a White House accusation that China was conducting industrial-scale efforts to copy and steal American frontier technology, placing the model preview in a wider U.S.-China technology tension context.

Terms & Concepts
  • DeepSeek-V4-Pro: A version of DeepSeek’s V4 model line included in the open-source preview, which DeepSeek’s API page describes as currently having limited throughput.
  • Huawei Ascend NPUs: Huawei’s AI-focused neural processing units; DeepSeek said its V4 preview was optimized for the Ascend platform and official documents said its expert parallel scheme was validated on these chips.
  • UMA dispute rulings: A dispute-resolution process used for prediction-market settlement when participants challenge whether an event outcome satisfies the market’s formal criteria.