Daedalus-150M: CPU-first hybrid tops data-heavier baselines

"Small language models are usually built like large ones and then squeezed onto a CPU afterwards," Christos Koutsiaris writes in the abstract of Daedalus-150M, an arXiv preprint dated August 20, 2026. "We did the opposite: we fixed the target first, one user, one token at a time, 4-bit weights, ordinary CPU, and chose the architecture to suit it."

The result is a 150-million-parameter language model with full attention in only 6 of its 18 blocks. The other 12 blocks run short convolutions with a fixed two-timestep memory width, so per-token memory does not grow with context length. Trained on 59.9 billion tokens, it reports a validation bits-per-byte of 0.8685 and scores 47.31 on a five-task benchmark against a 42.20 baseline.

Koutsiaris reports the model outperforms GPT-2 124M, Pythia-160M, OPT-125M and GPT-neo-125M despite each of those seeing three to six times more training data, and exceeds MobileLLM-125M, which was trained on roughly a trillion tokens. He also claims 1.76x faster decoding at 2048 tokens of context versus a comparable all-attention model, and a 4-bit file 6.3% smaller.

The paper is a single-author preprint and the abstract does not enumerate the five benchmark tasks or specify the CPU used for the timing.