Research paper published introducing FlashMemory-DeepSeek-V4 with Lookahead Sparse Attention for ultra-long context

DeepSeek V4 technical disclosure: 500K context with 90% less KV cache using FlashMemory

DeepSeek v4 API pricing permanently cut 75% to $0.43/M input tokens and $0.87/M output tokens

DeepSeek v4 launched, exposing performance gap in AMD ROCm vs NVIDIA CUDA

Achieved top score on METR time horizon benchmark, handling 90-minute software tasks

Deployment of DeepSeek V4 on Huawei's Ascend AI chips discussed as part of China's semiconductor counterstrategy.

Began limited gray-scale testing with a new tiered interface (Fast, Expert, Vision modes).

Achieved state-of-the-art status on most benchmarks according to preliminary evaluations