Agens Volundr 32B Preview

Agens Volundr is a dense 32B open-weights model engineered for extensive working sessions and complex operational tasks. Key features include: * Optimized KV cache management for long contexts * High token decode rate at extended context lengths * Broad context window (262K) * Open-source Apache-2.0 license * Hardware-efficient operation with BF16 and INT4 support This model is designed to sustain performance over prolonged interactions. By utilizing linear attention with a fixed-size state for 54 of its 72 layers, only 18 layers maintain a KV cache, making long context processing significantly more affordable on standard hardware setups. This innovative architecture ensures that decode speeds remain high, even at very large context windows, achieving 23.9 tokens per second at 128K context using two 48 GB GPUs. Agens Volundr supports a 262K context window, offering expansive capacity for intricate projects and detailed information processing. It is released under an Apache-2.0 license, providing flexibility for various applications. The model runs efficiently on Blockway's open sglang build, with planned support for GGUF, making it accessible for a wider range of users. Ideal for developers and organizations building sophisticated operational frameworks, this model is perfect for applications requiring deep contextual understanding and sustained performance over long periods. It offers a powerful foundation for developing advanced systems that demand both capacity and efficiency.