DeepSeek releases V4 Pro and Flash models with 1M-token context and MIT license, positioning Chinese AI labs as credible open-weight alternatives to closed frontier models

DeepSeek published two new AI models on April 23–24, 2026: V4 Pro (1.6 trillion total parameters with 49 billion active) and V4 Flash (284 billion total with 13 billion active). Both support 1 million token context (8× longer than V3.2), include reasoning modes, and are released under an MIT open-source license with a 58-page technical report. Pricing: V4 Pro at $1.74/$3.48 per 1 million input/output tokens, V4 Flash at $0.14/$0.28.

The new long-context architecture uses a hybrid attention system with shared key-value vectors and compressed token streams that reduce memory footprint to 9.62 GB per sequence (versus 83.9 GB for V3.2)—nearly 9× smaller. This means V4 Pro can run on a single GPU node (8× B200s) and processes 150+ requests per second on Blackwell Ultra hardware, making real-time multi-turn agent work (like autonomous coding assistants) practical on enterprise infrastructure.

Independent benchmarks place V4 Pro as the #2 open-weight model globally, ahead of Kimi K2.6 on agentic real-world tasks (1554 vs. 1484 score) and strongest on long-context and coding, though still behind GPT-5.x and Claude Opus 4.7 on overall capability. For software engineers and researchers using open models, this closes the gap to proprietary frontier models for long-document reasoning and code generation—previously a major weakness of open alternatives.

V4 models are immediately available via Huawei Ascend chip compatibility (China's export-controlled alternative to NVIDIA GPUs) and deployed by third-party API providers (Together, Baseten, Nous Research) within hours. DeepSeek signaled that V4 Pro pricing could drop significantly once Huawei Ascend 950 clusters scale in H2 2026, making frontier-class reasoning tasks cheaper for startups and researchers without NVIDIA access.