One year after DeepSeek’s R1 release shocked Silicon Valley, the Chinese AI lab is back with a pair of models that edge the open-source world closer than ever to closed frontier systems. DeepSeek V4 Pro and V4 Flash, released in preview on April 24, 2026, represent the most ambitious open-weight models the company has shipped — and the pricing alone is set to rattle enterprise AI procurement teams.
What DeepSeek V4 Actually Is
V4 Pro is a mixture-of-experts architecture with 1.6 trillion total parameters and 49 billion active parameters per inference pass — making it the largest open-weight model publicly available as of this writing. V4 Flash takes a leaner path: 284 billion total parameters, 13 billion active, optimised for speed and cost at the expense of raw capability ceiling.
Both models share a 1 million-token context window, a leap that allows developers to send entire codebases, lengthy regulatory documents, or multi-session conversation histories as a single prompt. DeepSeek attributes much of the context-retention improvement to a proprietary technique it calls Hybrid Attention Architecture, which it says reduces degradation in long-sequence tasks compared to standard transformer attention mechanisms.
Pricing is where the headlines are. V4 Flash costs $0.14 per million input tokens and $0.28 per million output tokens. V4 Pro comes in at $0.145 per million input and $3.48 per million output. For context, OpenAI’s comparable frontier offering prices output tokens at roughly an order of magnitude higher.
Performance: Close, but Not There Yet
DeepSeek’s internal benchmarks claim V4 closes the gap materially on reasoning and coding evaluations. Independent testing cited by TechCrunch and Simon Willison’s analysis suggest the models are competitive with current-generation frontier systems on mathematical reasoning and code generation, but trail on factual knowledge tests — specifically lagging behind OpenAI’s GPT-5.4 and Google’s Gemini 3.1 Pro.
Analysts describe the gap as approximately three to six months of development trajectory. That is a meaningful difference for enterprise customers in regulated sectors, but it suggests the competitive ceiling for open-source AI continues to rise at a pace that makes closed-model pricing increasingly hard to justify for cost-sensitive workloads.
Strategic Context: Chips, Cost, and China
Fortune reports that V4 integrates closely with Huawei’s Ascend AI chip ecosystem, a signal of China’s accelerating effort to build an AI stack independent of NVIDIA. With US export controls tightening access to high-end NVIDIA silicon in China, DeepSeek’s ability to release competitive models trained on domestically available hardware is itself a benchmark result — one with geopolitical weight beyond any coding leaderboard.
For Western AI developers and enterprise buyers, DeepSeek V4’s arrival continues the now-familiar pattern: China ships competitive open-source capability at pricing that compresses margins for API-first AI businesses. Whether V4 Pro ultimately displaces frontier models in production workloads depends on whether the knowledge benchmark gap matters for a given use case. For coding assistants, document processing, and agentic pipelines running over long contexts, it very well may not.
V4 Pro and V4 Flash are available today via the DeepSeek API. Open weights are expected to follow the preview period under the same permissive licence as prior DeepSeek releases.
Sources: TechCrunch (April 24, 2026), Fortune (April 24, 2026), DeepSeek API Docs, Simon Willison’s analysis, Al Jazeera (April 24, 2026)
Discussion
Sign in to join the discussion.
No comments yet. Be the first to share your thoughts.