Agens Volundr 32B Preview

Agens Volundr 32B Preview is an open-source AI large language model built for long-context coding, tool use, and extended agent sessions on self-hosted hardware. It is a dense 32-billion-parameter model (no mixture-of-experts) released under the Apache-2.0 license, with a 262K-token context window and a memory design that keeps the resource footprint predictable. Instead of letting the KV cache grow without limit, 54 of its 72 layers use a fixed-size state, and it adds Blockway Compressed-Sparse Attention for selective long-range token access plus an Engram n-gram memory that lives in host RAM rather than GPU memory. The practical result is that it runs on modest hardware: the BF16 build fits on two 48GB GPUs or a single H200, while INT4 quantization drops it to 31.7 GiB and runs on one 48GB GPU, sustaining about 24 tokens/second decode from 8K up to 128K context. For teams that want a commercial-grade model they can run in-house without cloud dependencies, this is a concrete option, and weights are available now on Hugging Face. BestAIFor readers evaluating self-hosted LLMs will find the fixed-state layer approach the main differentiator to test against the context length they actually need.

Self-hosting an open-source LLM for long-context coding, tool use, and extended agent sessions without cloud dependencies