According to CoinMeta, a British research institution AI in collaboration with researchers from the Swiss Federal Institute of Technology Zurich have open-sourced Latent Sparse Attention ( LSA ), a sparse attention mechanism designed for ultra-long contexts. Models such as DeepSeek, MiniMax, and Zhipu have already adopted sparse attention to process only a small number of key token at a time in order to reduce computation costs. In experiments involving 10 million token of GLM-5.3-Flash, researchers found that 98% of the pre-filling time for single-layer sparse attention was used for information filtering. LSA grouped token based on content similarity, and then used an indexer to score these groups to select relevant token. The researchers integrated this approach into GLM-5.3-Flash and Qwen-3.8-27B, retaining the original model weights while additionally training a small number of indexing parameters. On NVIDIA H200, the single-layer pre-filling of GLM was 22.7 times faster than the original sparse attention method, and the overall model's initial token latency was reduced by about 55%. Furthermore, the single-layer pre-filling of Qwen was 72.4 times faster, and the overall model's initial token latency was reduced by about 89%.