Emerging Technology — 2026-04-26
DeepSeek Unveils V4 Flash and Pro Models, Challenging Frontier AI Rivals at Fraction of Cost
DeepSeek launched preview versions of V4-Pro and V4-Flash on April 24. V4-Pro has 1.6 trillion total and 49 billion active parameters, described as the biggest open-weight model available. V4-Flash has 284 billion total and 13 billion active parameters. Both are mixture-of-experts architecture with 1 million token context windows, supporting dual Thinking and Non-Thinking modes. Pricing undercuts competitors with V4-Flash at 0.14 and 0.28 per million input and output tokens. DeepSeek claims V4-Pro outperforms OpenAI GPT-5.2 and Google Gemini 3.0 Pro on some reasoning benchmarks, matches GPT-5.4 in coding competitions, and approaches Claude Opus 4.5 on agentic tasks. Both are text-only open-weight models available via API and Hugging Face.
Analysis
Export controls on Western model access make V4-Pro a capability floor Chinese industry can build on without license exposure, and adoption by at least three major Chinese technology firms islikely within 60 days given open-weight availability, aggressive pricing, and regulatory necessity. The self-reported benchmark claims against GPT-5.2 and Claude Opus 4.5 have not been independently replicated, and prior Chinese frontier launches have consistently seen curated benchmarks outperform real-world deployment; the capability-gap narrowing at the frontier model tier remains conditional. V4-Flash at $0.14 per million input tokens disrupts inference economics independent of how those benchmarks resolve.
Export controls on Western model access make V4-Pro a capability floor Chinese industry can build on without license exposure, and adoption by at least three major Chinese technology firms is