Volundr’s new release carries conversation state between agent turns
Blockway has released new Agens Volundr weights aimed at extended sessions with AI agents. A continuation feature reuses the state from earlier input when a conversation resumes, leaving subsequent requests to process the added messages. The design uses both growing memory caches and fixed-size layer states. Adapted serving software is required, and the separate inference engine initially supports text alone.
Artificial Intelligence··Evening
Session continuation keeps earlier conversation state
The model developer Blockway has released Agens Volundr 32B v1 with a session-continuation feature for long AI agent conversations. The company’s model card describes carrying the state associated with an earlier conversation prefix into the next request. Newly appended messages receive processing on subsequent turns, while the state from earlier input is reused. The release follows the October 5 Preview and includes completed continued pre-training and additional training aimed at extended agent use.[1]
Different layers hold different kinds of memory
Volundr combines layers that retain a growing key-value cache with layers holding a fixed-size state. A key-value cache saves information from earlier input for reuse. Of the model’s 72 layers, 18 retain that growing cache while 54 use fixed-size states. The main model accepts images as well as text. Deployment depends on Blockway’s adapted serving software, making the software path part of using the released weights.[1]
The separate engine initially serves text
Blockway also provides the MIT-licensed Agens Engine for inference, the process of running a trained model to produce output. Its first release supports text only. The engine can move some data into host memory and store weights at different bit widths. The model card separately reports company-run evaluations of repetition and task completion, including some regressions against the Preview on general evaluations. Some tables select the best of several runs; those figures come from the vendor’s testing.[1]
