Brick: Spatial Capability Routing for the Mixture-of-Models (MoM) Paradigm

A new router called Brick outperforms the best single large language model and all tested routing systems by scoring models across six capability dimensions and factoring in per-query difficulty, according to a paper posted to arXiv [1]. Existing LLM routers rely on surface features such as domain labels, keywords, and token count, ignoring the within-domain variance that determines model success [1]. Frontier models cost ten to one hundred times more than local open-weight models, so at production scale even small per-request savings become a direct cloud-bill lever [1]. The authors present Brick, a multimodal router that scores each model on six capability dimensions, combines this with a per-query difficulty estimate, and dispatches via a cost-penalized geometric rule [1]. A continuous preference knob lets operators slide between max-quality and max-saving profiles at deploy time [1]. On a benchmark of 5,504 queries, Brick at max-quality reaches 76.98% accuracy, beating the best single model at 75.02% and all tested routers [1]. At a neutral cost-quality profile, Brick achieves 74.11% accuracy at 4.71x lower cost than always using the strongest model [1]. At min-cost, it cuts cost 22.15x with 11.85 points accuracy loss [1]. Median latency drops from 51.2 seconds to 22.8 seconds [1]. The work lands as the economics of large language model deployment remain under scrutiny. DeepSeek, the Chinese AI company, trained its V3 model for a reported US$6 million, far less than the US$100 million cost for OpenAI's GPT-4 in 2023, using approximately one-tenth the computing power consumed by Meta's comparable Llama 3.1 model [6]. DeepSeek's models are described as open-weight, meaning the exact parameters are openly shared but the training data is not openly licensed [6]. The company's success against larger rivals has been described as "upending AI" [6]. The Brick paper appears on arXiv, which since 2022 has integrated with Hugging Face Spaces to embed interactive demos directly alongside papers [3][4]. The integration allows users to try state-of-the-art machine learning research without writing any code [5]. Demos are built using open-source tools such as Gradio and Streamlit and leverage models and datasets available on the Hugging Face Hub [4]. Large language models are a type of machine learning model designed for natural language processing tasks such as language generation, trained with self-supervised learning on vast amounts of text [7]. The field has drawn researchers including Douwe Kiela, who co-authored the foundational paper on retrieval-augmented generation at Meta AI in 2020 and later served as Head of Research at Hugging Face [8].