Subquadratic launches 12M-token LLM, claims major architectural shift

Subquadratic, a Miami-based startup, has emerged from stealth claiming to have developed the first Large Language Model (LLM) that does not utilize quadratic attention. This architectural innovation reportedly enables the model to process a context window of 12 million tokens at a significantly reduced cost compared to existing frontier models. AI

Summary written by gemini-2.5-flash-lite from 1 source. How we write summaries →

IMPACT Potential to drastically lower inference costs for LLMs with extremely long context windows.

RANK_REASON Startup announces a novel LLM architecture with a large context window and reduced cost. [lever_c_demoted from significant: ic=1 ai=1.0]

Read on Mastodon — fosstodon.org →

• Subquadratic
• Miami
• LLM
• transformer

• model release
• paper