Standard OCR decoders grow KV cache linearly with output length. Baidu’s R-SWA caps cache at Lm + min(n, T) ≤ Lm + n where n defaults to 128. Memory and per-step latency stay flat as document length grows.
MIT license, open weights, dual Transformers and SGLang support. vLLM integration added post-release. Only 500M active parameters keep inference cheap. R-SWA gains arrive without measured accuracy cost on single pages per technical report.