Released Rnj-1.5, a long-context extension of Rnj-1 scaling to 160k tokens! By moving from all-global to block-local/global attention layers, our 8B model hits 40% on SWE-bench Verified (thread).