DeepSeek V4 Pro is a large-scale Mixture-of-Experts model that pairs 1.6 trillion total parameters with 49 billion activated per token, sharing its underlying architecture with the lighter DeepSeek V4 Flash sibling and introducing a hybrid attention system aimed at efficient long-context processing. It exposes a three-tier reasoning control surface, with Non-think, Think High, and Think Max effort modes, allowing callers to trade latency for depth on the same checkpoint. This design positions V4 Pro as the frontier-grade option in the V4 lineup, while Flash is recommended for high-throughput agents and routine reasoning where a smaller active footprint is acceptable.

In benchmark results reported by FriendliAI with Think Max enabled, V4 Pro led the open-weight field on LiveCodeBench at 93.5, Codeforces at 3206, and SWE Verified at 80.6, reinforcing its stated focus on software engineering and multi-step problem solving. The model is offered through OpenRouter's unified API alongside other DeepSeek releases and is also deployable on FriendliAI Dedicated Endpoints, making it practical for full-codebase analysis, multi-step automation, and large-scale information synthesis where both reasoning depth and long-context efficiency matter.