AnalysisAI ModelsNovember 17, 2026

RecSA brings million-token context at logarithmic cost

Read original source →haoailab.com

RecSA builds a tree over the KV cache and searches it level by level, cutting selection cost to O(log L). Trained only at 32K from Qwen3-4B on 0.62B tokens, it scores 60.5 on RULER at 1M versus 40.0 for Dense + DCA, and runs 3.2x faster prefill and 3.8x faster decoding at 524K on one H200.

1 source

More stories today

Open the live feed