A startup claims it broke through a bottleneck that’s holding back LLMs

▤ Summary

Subquadratic says it has built a new large language model, SubQ, that uses sparse attention to reduce the compute cost of long-context AI workloads. The company now backs up its claims with third-party testing from Appen, which found strong speed and retrieval results. The model appears especially relevant for coding and document-heavy tasks, where it can process far more text than many mainstream models. Even so, the article notes that skepticism remains because the evidence is still limited and the model is not widely available. The piece also points out that Subquadratic reused Qwen weights, which weakens the company’s claim that it has fully reinvented LLM architecture.

▦ Linked Companies

Subquadratic Inc.

$10M to $25M