First, to join the early access queue, you must submit a form on their website. https://subq.ai/
The startup Subquadratic, founded by ex-DeepMind and Meta engineers, claims to have developed an architecture that reduces processing costs by up to 1,000x compared to current models.
Here is the breakdown of the technical claims:
The bottleneck
Current LLMs face a scaling wall. Doubling the input data typically causes computational costs to explode exponentially. This inefficiency is the primary barrier to expanding context windows and model capabilities according to them
The Linear Solution *
Subquadratic’s model promises linear scaling. In this framework, doubling the data only doubles the processing requirement. They are reporting a 12-million-token context window, claiming a 52x efficiency gain at the 1-million-token scale compared to standard
Transformer architectures and the Impact on RAG.
If models can natively handle this much data without performance degradation, current workarounds like RAG and complex vector database pipelines could become obsolete. The model would simply process the entire dataset within the prompt.
The Reality Check, benchmarks, weights and etc...
The scientific community is currently calling for peer reviews. We have seen many "breakthroughs" fail to move past the whitepaper stage due to hardware constraints or hidden trade-offs in accurac
What is not a breakthrough here:
While the ex-DeepMind and Meta make those claims to attract venture capital, crucial technical limitations are being conveniently ignored by the startup, including the fundamental mathematical trade-off between simple data retrieval and complex global reasoning, the stark physical reality of hardware memory bandwidth bottlenecks that software alone simply cannot fix, and the glaring lack of independent peer review to verify whether this closed-source model is an actual architectural paradigm shift or just another heavily lossy, hybrid trick disguised as the next leap forward in artificial intelligence.
Subquadratic just pulled in a heavy $29 million in seed funding, backed by players like Vision Fund, Tinder’s co-founder, and early investors from OpenAI and Anthropic.
According to the website The New Stack, the company's valuation reached US$500 million.