DeepSeek-V4-Pro
The professional edition of the fourth-generation V4 series of large language models released by DeepSeek. In LSCBench it achieves a functional pass rate of 73.6% and a safety pass rate of 18.0%.
Timeline (last 90 days)
The top five in MMLU-Pro are separated by only 0.6 percentage points: Darwin-180B-RSI 88.1%, MiniMax-M2.1 88.0%, Intern-S2-Preview 88.0%, Qwen3.5-397B-A17B 87.8%, DeepSeek-V4-Pro 87.5%.
After the researchers added a sentence to four models requiring them to implement "ensuring security while meeting functional requirements," GPT-5.6-Sol's functional + security pass rate rose from 24.7% to 46.2%, GLM-5.3 from 21.3% to 38.2%, Qwen3.8-Max from 22.0% to 37.6%, and DeepSeek-V4-Pro from 18.0% to 19.1%.
DeepSeek-V4-Pro's 1.6T MoE requires the 1,440 GB memory pool of 8×B200 Blackwell under FP8.
The unified Radix Cache added an Agentic-Aware tail-optimized LRU eviction strategy, which reduced p99 ITL by 43.9% at concurrency 32 in an agentic trajectory test on DeepSeek-V4-Pro FP4 with 4-card B300.
Through request-owned speculative KV, in a DeepSeek-V4-Pro 1P1D, MI355X, concurrency-256 scenario, the average TTFT was reduced from 19.4 seconds to 15.1 seconds.
DeepSeek-V4-Flash and DeepSeek-V4-Pro are priced 50% lower outside the hours of 12:00 to 18:00 UTC on weekdays and all day on weekends.
The original two-step allgather+scatter operation in TP LMHead has been merged into a single all-to-all communication, reducing LMHead latency from 320μs to 169μs and improving TPOT from 36.97ms to 35.67ms in DeepSeek-V4-Pro B200 decoding tests.
RisenX默认运行在DeepSeek-V4-Flash上，输入/pro命令可在下一轮切换到DeepSeek-V4-Pro随后自动切回，运行/preset max可整个会话使用Pro
All Facts (9)
V4-Pro的总参数与活跃参数比例约为32:1的稀疏比
90%UnverifiedV4-Pro uses a Mixture of Experts (MoE) architecture with a total parameter count of 1.6 trillion (1.6T) and 49 billion (49B) active parameters.
70%UnverifiedThe top five in MMLU-Pro are separated by only 0.6 percentage points: Darwin-180B-RSI 88.1%, MiniMax-M2.1 88.0%, Intern-S2-Preview 88.0%, Qwen3.5-397B-A17B 87.8%, DeepSeek-V4-Pro 87.5%.
50%UnverifiedAfter the researchers added a sentence to four models requiring them to implement "ensuring security while meeting functional requirements," GPT-5.6-Sol's functional + security pass rate rose from 24.7% to 46.2%, GLM-5.3 from 21.3% to 38.2%, Qwen3.8-Max from 22.0% to 37.6%, and DeepSeek-V4-Pro from 18.0% to 19.1%.
50%UnverifiedDeepSeek-V4-Pro's 1.6T MoE requires the 1,440 GB memory pool of 8×B200 Blackwell under FP8.
50%UnverifiedThrough request-owned speculative KV, in a DeepSeek-V4-Pro 1P1D, MI355X, concurrency-256 scenario, the average TTFT was reduced from 19.4 seconds to 15.1 seconds.
50%UnverifiedThe unified Radix Cache added an Agentic-Aware tail-optimized LRU eviction strategy, which reduced p99 ITL by 43.9% at concurrency 32 in an agentic trajectory test on DeepSeek-V4-Pro FP4 with 4-card B300.
50%UnverifiedThe original two-step allgather+scatter operation in TP LMHead has been merged into a single all-to-all communication, reducing LMHead latency from 320μs to 169μs and improving TPOT from 36.97ms to 35.67ms in DeepSeek-V4-Pro B200 decoding tests.
50%UnverifiedRisenX默认运行在DeepSeek-V4-Flash上，输入/pro命令可在下一轮切换到DeepSeek-V4-Pro随后自动切回，运行/preset max可整个会话使用Pro
50%