An arXiv preprint reports that Daedalus-150M, a roughly 160-million-parameter language model mixing full-attention and short-convolution blocks, generated text faster on a CPU than a similarly sized all-attention twin, with the difference widening as the prompt grew longer. The head-to-head was a computational comparison using one training seed.
At an empty context, the hybrid-to-twin throughput ratio was 1.20×. It rose to 1.45× at 512 tokens and 1.76× at 2,048 tokens, the model’s stated context length. The tests generated 128 tokens with 4-bit weights on eight threads; results were reported as means with standard deviations, but the number of repetitions was not given.
The advantage grew with context
The pattern also appeared against the external Peer-135M checkpoint: the hybrid’s throughput ratio was 1.06× at empty context, 1.49× at 512 tokens and 2.08× at 2,048 tokens. Because this was an external comparison rather than the matched twin test, it does not isolate the architecture as cleanly.
A simple bandwidth calculation predicted a 17% advantage at 2,048 tokens, far below the measured 76%. The paper argues that bandwidth alone cannot explain the result and points to cache-related latency and fixed costs for each layer as possible contributors. That explanation is an interpretation of the measurements, not a randomized causal test.
Quality stayed close, but the hybrid did not win every test
The speed result was not accompanied by a clear quality loss in the main head-to-head. On the pre-registered validation measure, the hybrid performed 0.81% better than the dense twin and cleared the study’s 0.5% margin floor.
On the five-task mean, however, the dense model was nominally ahead: 44.82 points versus 44.68 for the hybrid. The paper describes that as a downstream-quality match rather than a hybrid win; per-task winners were split, and the difference was about 0.24 standard deviations against roughly 0.58 standard deviations of suite variation.
In the full-model benchmark, Daedalus-150M posted a five-task mean of 47.31, clearing a fixed 42.20 bar by 5.11 points. Peer-135M scored 51.2, leaving it 3.9 points ahead, so the hybrid’s CPU result did not amount to a broader benchmark lead.
The main run used a ten-source English mixture with 16.93 billion unique tokens and a 59.9 billion-token budget. On the 645-million-token holdout, its validation bits-per-byte was 0.8685, compared with 0.9104 for the same architecture after 5 billion training tokens; the paper reports a 4.6% improvement associated with the remaining 55 billion tokens.
A close comparison with a meaningful caveat
Daedalus-150M has 160.49 million parameters in 18 blocks: six use full attention and 12 use short convolutions. Its model width is 768, and its stated context length is 2,048 tokens.
The controlled twin had 161.25 million parameters and 24 all-attention layers. Both models used identical data and training schedules for 5 billion tokens, were evaluated on a 645-million-token holdout, and were judged under a rule fixed before scoring. The parameter counts were within 0.5%, but the different layer layouts mean the comparison did not isolate architecture completely.
The hybrid’s 4-bit file was 95.56 MiB, compared with 101.62 MiB for the dense twin—a reported 6.3% reduction. Its effective storage was 4.99 bits per weight versus 5.29 for the twin.
The compact release still has engineering costs
The 4-bit format also showed a reported quality gap on held-out encyclopedic text. Perplexity, a measure of prediction error, was 9.75 with Q4_0 weights versus 9.18 at half precision—roughly a 6% difference, compared with 2.5% at the smaller 5-billion-token scale. A planned final 5% quantisation-aware training pass produced a non-finite loss on its first step, was disabled, and left the released model quantised after training.
Training also left many inactive channels in the short-convolution blocks. The dead-channel fraction was 47.928% at step 9,896 and 47.993% at step 30,041, corresponding to roughly 13.6 million inert parameters and 8.5% parameter inefficiency. Export pruning failed when the reference runtime rejected narrowed tensor shapes.
A result tied to one deployment setup
The evidence is narrow. The matched ablation used one seed and reported no confidence intervals, formal hypothesis tests or power calculation. The five-task suite is only a surrogate for broader language ability, and the model was evaluated as an English system.
The decode measurements used eight threads, excluded prompt processing and were affected by background machine load; the exact CPU was not reported. The study demonstrated the measured pattern only through the trained 2,048-token context and did not run retrieval-sensitive evaluations.
The findings are most relevant to single-user CPU generation at the tested context range. They do not establish that the same advantage will hold across other hardware, runtimes, thread counts, batch sizes, languages or workloads. Multi-seed tests, longer contexts and retrieval-sensitive evaluations remain open questions.
Paper data and sources
Original title: Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference
Authors: Christos Koutsiaris
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text