sparse attention
2 published articles
AI5 min read
Deep Tech
One routing pass to rule them all: 7x faster LLM decoding without the quality trade-off
CLSA uses one routing pass per sequence instead of one per layer, cutting KV-cache overhead while keeping token-level selectivity. Benchmarks show 17.1x throughput improvement at 128K context.
2026-07-26
Labs & ResearchFeatured3 min read
AI Labs & Research
MiniMax's M3 just wrote its own CUDA kernel, and opened the code
MiniMax M3 scores 83.5 on BrowseComp, edges past Opus 4.7, and handles up to 1M tokens natively. In a remarkable autonomy test, it self-optimized a GPU kernel from 7.6% to 71.3% peak utilization without human intervention.
2026-07-09