Long Offline Cross-Reference

You are building an offline model over 4000-token documents where answers often depend on matching phrases across distant sections. Latency is less important than capturing those long-range interactions. Which architecture should dominate the shortlist?