Full Breakdown
DeepSeek Unveils V4 AI Model Amid US-China Tech Tensions
4/24/2026, 8:35:11 PM
DeepSeek V4 Launch
On 24 April 2026 DeepSeek, a Hangzhou AI startup, released preview versions of its V4 large-language model in two variants—Pro and Flash. The launch follows OpenAI’s GPT-5.5 release a day earlier and builds on DeepSeek’s 2025 R1 model, which had matched leading U.S. systems at low cost.
Technical Specs and Pricing
V4-Pro has 1.6 trillion parameters; V4-Flash 284 billion. Both support a one-million-token context window, ten times DeepSeek’s prior limit. DeepSeek prices output at $3.48 per million tokens for Pro and $0.28 for Flash, versus $30–$25 for OpenAI. The models run on Huawei Ascend 950 chips; training hardware was not disclosed.
Market Impact and AI Race
The V4 launch underscores China’s drive for AI self-reliance by pairing open-source software with domestic silicon, cutting reliance on U.S. Nvidia GPUs. Lower inference costs and a one-million-token context may expand AI use in finance, healthcare and research, especially where budgets are limited. SMIC and Hua Hong shares rose 9% and 15%, while MiniMax and Zhipu fell 7–9%.
Official Statements & Responses
DeepSeek said V4 offers “world-leading long-context with drastically reduced compute and memory costs” and is open-source. Huawei said its Ascend 950 supernode clusters support V4 inference. Michael Kratsios warned of “industrial-scale distillation campaigns” by Chinese entities. Premier Li Qiang described China’s AI models as “spearheading AI ecosystem,” and Foreign Ministry spokesperson Guo Jiakun called the U.S. claims “groundless.”
Criticism & Opposition
OpenAI and Anthropic have accused DeepSeek of illicitly extracting capabilities from their proprietary models through distillation, a claim the company denies. Analyst Ivan Su described V4 as a “competent follow-up” that falls short of the breakthrough impact of the earlier R1 release and called for independent testing.
Conflicting Reports & Gaps
DeepSeek did not confirm whether Nvidia GPUs were used for V4 training, while Huawei said its Ascend chips were employed for part of the Flash model’s training. Performance claims rely on DeepSeek’s internal benchmarks; no third-party evaluations have been released.
Verbatim Quotes
- “Based on the benchmark results, it does appear DeepSeek V4 is going to be very competitive against its U.S. rivals,” — Lian Jye Su, chief analyst, Omdia
- “pivotal milestone for China’s AI industry,” — Marina Zhang, associate professor, University of Technology Sydney
- “DeepSeek's V4 preview is a serious flex,” — Neil Shah, vice president of research, Counterpoint Research
- “The US has evidence that foreign entities, primarily in China, are running industrial-scale distillation campaigns to steal American AI,” — Michael Kratsios, chief science and technology adviser to President Donald Trump
What’s Next
DeepSeek expects V4-Pro pricing to fall after large-scale deployment of Huawei Ascend 950 supernodes later in 2026, and it is pursuing a funding round that could value the firm above $20 billion, with potential stakes from Alibaba and Tencent. The U.S.–China summit slated for May may address the broader dispute over AI intellectual-property theft.
