1.65T checkpoint / 49B active target model (MoE; includes attached DSpark module)
1,048,576 tokens
MIT
Official DeepSeek-V4-Pro release superseding the April preview, with substantially enhanced agentic capabilities and especially large reported gains on production-oriented tasks. The checkpoint builds on the V4-Pro target architecture and bundles an attached DSpark speculative-decoding model in the same repository. Its published configuration describes a 61-layer DeepSeek-V4 causal architecture with 384 routed experts plus one shared expert, six routed experts selected per token, a 7,168-dimensional hidden state, 128 query-attention heads, one key/value head, compressed sparse-attention components, FP4 expert weights, FP8 quantization for other supported components, and YaRN scaling from an original 65,536-token window to 1,048,576 tokens. DeepSeek reports 87.9 on Terminal Bench 2.1, 61.5 on NL2Repo, 83.3 on Cybergym, 62.7 on DeepSWE, 74.1 on Toolathlon-Verified, 25.7 on Agents' Last Exam, and 31.8 on AutomationBench Public. The model also reaches 42.7 on Humanity's Last Exam without tools and 60.0 with tools, plus 71.1 on DSBench-FullStack and 67.2 on DSBench-Hard in DeepSeek's internal evaluations. It supports low, high, and max reasoning-effort settings, uses DeepSeek's custom OpenAI-compatible message encoding rather than a Jinja chat template, and includes local deployment support for vLLM, SGLang, and DeepSeek's reference inference implementation. DeepSeek recommends temperature 1.0 and top-p 0.95 for agent tasks and permits output limits up to 384K tokens for high and max reasoning effort.