DeepSeek V4

DeepSeek V4 is the 2026 successor to the V3 line, released as previews on April 24, 2026 in two sizes: V4-Pro, a 1.6-trillion-parameter MoE with 49B active per token, and V4-Flash, a 284B MoE with 13B active. Both are distributed under the MIT license with weights on Hugging Face and ModelScope, making V4-Pro one of the most capable open-weights models available at the time of writing.

The headline feature is long-context efficiency. A hybrid attention scheme combining compressed sparse attention and heavily compressed attention lets the model competently handle a 1M-token window while using a fraction of the inference FLOPs and KV-cache memory of the V3 generation. DeepSeek's own reporting puts V4-Pro's 1M-context inference cost at about 27% of V3.2's. On benchmarks, the vendor claims 93.5 on LiveCodeBench, a 3206 Codeforces rating, and 80.6 on SWE-bench. V4 is text-only and aimed squarely at coding agents and long-horizon agentic workflows.