Runs on ≤32 GB · Hugging Face ↗
Not independently scored by Artificial Analysis
Released 2026-10-03 by Blockway, Apache-2.0, as a Preview: the full v1 will replace these weights. Own VolundrForConditionalGeneration architecture, 72 layers: 54 of KDA linear attention, 17 of Blockway's compressed-sparse attention and one dense, plus Engram lookup tables. The card names no base model, but the vocabulary (248,320) and the dimensions match Qwen3.5-27B.
Real GGUF file sizes = weight VRAM. Add the KV cache for your context (≈3 GB at 32K, fp16). Read from the GGUF header, but this architecture has never been checked against llama.cpp — treat it as an estimate.
| Quantization | Size |
|---|---|
| Q4_K_M | 26.6 GB |
| Q8_0 | 36.3 GB |
| BF16 | 64.5 GB |
Stock llama.cpp refuses the file with "unknown model architecture: volundr": it needs the `volundr` branch of github.com/BlockWayz/llama.cpp. The Q4_K_M is 26.6 GB because the KDA projections are kept at Q8_0, so it needs a 32 GB card; the 4.3 GB of Engram tables can stay in host RAM with --override-tensor "engram_embd=CPU". Text only in the GGUF.