@HuggingPapers
DeepSeek-V4 now runs 500K context with 90% less KV cache FlashMemory introduces Lookahead Sparse Attention: a tiny Neural Memory Indexer predicts which chunks your next tokens will need, keeping only 13.5% of cache in GPU memory—with zero backbone retraining and better accuracy. https://t.co/qxclBmFV3n