@ModelScope2022: 1M-token context with only ~3B parameters active per token. Meituan’s LongCat-Flash-Lite-Sparse brings sparse attention…

X AI KOLs Timeline Models

Summary

Meituan released LongCat-Flash-Lite-Sparse, a sparse-attention model supporting 1M-token context with only ~3B active parameters per token, achieving strong SWE-Bench scores under an MIT license.

1M-token context with only ~3B parameters active per token. Meituan’s LongCat-Flash-Lite-Sparse brings sparse attention to long-context agents. https://modelscope.ai/models/meituan-longcat/LongCat-Flash-Lite-Sparse… The sparse variant scores 68.2 on SWE-Bench Verified and 59.33 on SWE-Bench Multilingual, up from 54.4 and 38.1 for its dense predecessor. LongCat Sparse Attention combines streaming-aware, cross-layer, and hierarchical indexing to reduce long-context computation while retaining model quality. 69B total parameters. MIT License.
Original Article
View Cached Full Text

Cached at: 08/03/26, 01:43 PM

1M-token context with only ~3B parameters active per token. Meituan’s LongCat-Flash-Lite-Sparse brings sparse attention to long-context agents. https://modelscope.ai/models/meituan-longcat/LongCat-Flash-Lite-Sparse…

The sparse variant scores 68.2 on SWE-Bench Verified and 59.33 on SWE-Bench Multilingual, up from 54.4 and 38.1 for its dense predecessor. LongCat Sparse Attention combines streaming-aware, cross-layer, and hierarchical indexing to reduce long-context computation while retaining model quality. 69B total parameters. MIT License.

Similar Articles

MiniMax Sparse Attention

Hugging Face Daily Papers

MiniMax Sparse Attention introduces a blockwise sparse attention mechanism that achieves significant speedups for ultra-long-context LLMs, reducing per-token attention compute by 28.4x at 1M context with wall-clock speedups of 14.2x for prefill and 7.6x for decoding on H800 GPUs. The method is accompanied by an open-source inference kernel and a publicly released multimodal model.