CPU-first Daedalus-150M trades most attention for bounded-memory convolutions and reports length-dependent speedup

Daedalus-150M is a 150M-parameter model that keeps full attention in only 6 of 18 blocks and uses short convolutions with two-timestep memory in the rest, targeting 4-bit CPU inference from the start. The authors report a 0.81% quality win over an equally sized all-attention model trained on the same data, a 6.3% smaller 4-bit file, and a 1.76x decode speedup at 2048 tokens of context that grows with context length.

Hugging Face Daily Papers — Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference

August 25, 20261 source55% confidence

70Strategic

Strategic score

Evidence strength45

Technical novelty80

Commercial potential65

Daedalus-150M is an architecture-first attempt at CPU inference: the target (one user, one token, 4-bit weights, ordinary CPU) is fixed before the architecture. The model places full attention in only 6 of 18 blocks; the other 12 use short convolutions whose memory is two timesteps wide, so two-thirds of the network does not re-read a growing cache. Reported speed advantage is near zero at empty context and grows with length, which the authors identify as the mechanism's signature.¹

The evidence for the architecture is a controlled same-size, same-data comparison against an all-attention model, with a pre-committed quality metric: the hybrid won by 0.81%, matched downstream tasks, made a 6.3% smaller 4-bit file, and decoded 1.76x faster at 2048-token context. Against an external model of similar size it was 2.08x faster. The authors also report a bandwidth calculation predicting only 1.17x, indicating memory volume alone does not explain the gap.¹

Context

The model was trained from scratch on 59.9B tokens and scores 47.31 on a five-task benchmark against a bar of 42.20 fixed before training. It is compared to GPT-2 124M, Pythia-160M, OPT-125M, GPT-neo-125M (all trained on 3-6x more data) and MobileLLM-125M; validation bits-per-byte is 0.8685.¹

Assessment

What is new is not another smaller Transformer but an architecture whose memory footprint is bounded independently of sequence length in most of the network. The growing speed advantage with context is the load-bearing observation: a merely leaner model would produce a constant speedup, whereas this design predicts a length-dependent one. The 1.76x measured speedup exceeding the 1.17x bandwidth prediction suggests a qualitative difference, but the mechanism that produces the extra gain is not specified in the abstract.¹

Caveats

This is a single arXiv abstract, not peer-reviewed, and no code or weights are listed in the metadata. The five-task benchmark and quality metric are unnamed, the external model is unidentified, and the bandwidth calculation's assumptions are not shown. A change of mind would follow from a replication showing the speed advantage is constant rather than length-dependent, or from detailed profiling showing the convolution blocks are not the cause.¹

Source text

Hugging Face Daily Papers — Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference · August 20, 2026

Computer Science > Information Retrieval

Title:Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference

Abstract:Small language models are usually built like large ones and then squeezed onto a CPU afterwards. 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 keeps full attention in only 6 of its 18 blocks. The other 12 use short convolutions whose memory is two timesteps wide no matter how long the conversation gets, so two thirds of the network never re-reads a growing cache. Trained from scratch on 59.9B tokens, the model scores 47.31 on a five-task benchmark against a bar of 42.20 that was fixed before training began. It beats GPT-2 124M, Pythia-160M, OPT-125M and GPT-neo-125M, all trained on three to six times more data, and exceeds MobileLLM-125M's published score despite that model seeing a trillion tokens. Validation bits-per-byte is 0.8685. To check the architecture rather than the training recipe, we trained a conventional all-attention model of the same size on the same data, and wrote down the winning condition before scoring either. The hybrid won the chosen quality metric by 0.81%, matched it on downstream tasks, produced a 6.3% smaller 4-bit file, and decoded 1.76x faster at 2048 tokens of context, 2.08x against an external model of similar size. In every measurement the speed advantage is near zero at an empty context and grows with length, which is what the mechanism predicts and what a merely leaner model would not show. A simple bandwidth calculation predicts only 1.17x, so memory volume alone does not explain the gap. We also report what did not work: an unmitigated 4-bit quality cost, roughly half the convolution channels ending up inert and impossible to remove, and a vocabulary larger than this model size warrants.

Access Paper:

Current browse context:

Bookmark

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.

Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.

Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.

The source text above remains the property of its publisher and is reproduced for commentary.

Hugging Face Daily Papers — Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference

August 20, 2026endistinct angle

Daedalus-150M: A Convolution-Attention Hybrid Designed for CPU Inference

↑ Back to top