The Agens Volundr 32B v1 model resolves the repetition loops seen in its preview, dropping SWE‑bench loop occurrences from 36/50 to 0/50. It also improves reasoning continuity and session handling, enabled by an MIT‑licensed engine that supports up to 262K token context (measured at 239K) while staying within a 16 GB memory budget.