Agens Volundr 32B v1
RELEASED · v1Our hybrid-architecture model is out: about 32 billion parameters, with only 18 of its 72 layers keeping a KV cache, built for agent workloads and long-running sessions. It is derived from Qwen3.8-27B (Apache-2.0): a new 72-layer architecture transplanted onto its weights, then continued pre-training and agent post-training. The lineage is documented on the model card.
Released 18 Oct 2026 · Apache-2.0 · Hugging Face and ModelScope
Meet Agens Volundr 32B v1 →