Blockway releases preview of Agens-Volundr-32B model for reasoning, code and agentic work

The model Blockway/Agens-Volundr-32B-Preview with the text-generation task has been published on HuggingFace. This is a preview release from Blockway: the architecture, format and post-training pipeline are final, but continued pretraining is still ongoing, and the full v1 version will replace this checkpoint. The model is positioned for reasoning, coding and agentic work, speaks English, Simplified and Traditional Chinese and Cantonese, accepts text and images, and can call tools. The author notes that the weak spot in the preview is long agentic sessions.

What it means

The model contains 32.7 billion parameters, the weights are published in safetensors format, the repository totals 60.9 GB, and the largest file is 8 GB. The license is apache-2.0. It runs via transformers. There is one quantized GGUF build on the site: Q4_K_M weighing 24.8 GB — this is the smallest working build, not counting quants worse than 4 bits. Running the full model on your own hardware will require a substantial amount of VRAM, while the quantized Q4_K_M at 24.8 GB is a more realistic option for local use. We don't count builds worse than 4 bits. More on local AI tool news — in the CPU3D news feed.