Full Breakdown
DeepSeek Launches V4: Open-Source AI Model with One-Million-Token Context
4/25/2026, 9:13:17 AM
DeepSeek Unveils V4 Preview
On 24 April 2026, Hangzhou-based DeepSeek released preview versions of its V4 large-language model. The Pro variant holds 1.6 trillion total parameters; the Flash variant holds 284 billion. Both support a one-million-token context and are open-source.
From R1 Shock to Self-Sufficiency
DeepSeek first drew global attention in Jan 2025 with the R1 model, which matched U.S. systems at a fraction of the compute cost. The success sparked a “Sputnik moment” and heightened U.S.–China AI rivalry, especially after export bans on Nvidia chips.
Model Specs and Benchmarks
V4-Pro activates 49 billion parameters; V4-Flash activates 13 billion. DeepSeek says the Pro version leads open-source models in world-knowledge tests and trails only Google’s Gemini-Pro-3.1. Pricing is $3.48 per million output tokens for Pro and $0.28 for Flash, versus $30-$25 for OpenAI and Anthropic services.
Strategic Impact
Running on Huawei’s Ascend 950 chips, V4 reduces reliance on foreign hardware, a core goal of China’s AI self-sufficiency. SMIC shares rose 10 % after the launch, while rivals MiniMax and Zhipu fell 7-9 %. Lower costs could expand AI use in Chinese municipalities, healthcare and finance.
Official Statements & Responses
DeepSeek’s WeChat post said the one-million-token window will become “standard for all official services.” Huawei confirmed its Ascend SuperNode line “fully supports the DeepSeek V4 series.” Beijing’s foreign ministry called U.S. theft claims “baseless,” while Michael Kratsios warned of “industrial-scale distillation campaigns” against American AI.
Criticism & Opposition
U.S. officials and firms such as OpenAI and Anthropic allege DeepSeek illicitly “distilled” proprietary models for V4. Critics also note the chatbot’s refusal to answer politically sensitive queries, raising censorship and privacy concerns.
Conflicting Reports & Gaps
Sources disagree on how much Nvidia versus Huawei hardware was used for V4 training; DeepSeek has not disclosed chip mixes or costs. Benchmarks place V4-Pro 3-6 months behind top closed models, yet the firm claims “world-class” reasoning without independent verification.
Verbatim Quotes
- “The US has evidence that foreign entities, primarily in China, are running industrial-scale distillation campaigns to steal American AI,” — Michael Kratsios, U.S. Science & Technology Adviser
- “China-made large AI models spearheaded the development of the global open-source AI ecosystem,” — Li Qiang, Premier of China
- “This is a big deal for China’s AI industry,” — He Hui, Director, Omdia semiconductor research
- “R1 shocked US markets because no one expected a Chinese model to compete at that level. V4 is simply a follow-through on that same trend, and trends don’t make headlines the way shocks do,” — Ivan Su, Senior Equity Analyst, Morningstar
What’s Next
DeepSeek is pursuing a funding round with Tencent and Alibaba that could value the startup above $20 billion. Huawei plans to scale Ascend 950 supernode clusters in the second half of 2026, which DeepSeek says will slash Pro-service costs. The company also announced that its older models will be retired by July 2026, consolidating traffic onto V4.
