Miami-based KI startup Subquadratic has emerged from stealth with a bold claim: its new model SubQ runs up to 56 times faster than established LLMs while consuming significantly less energy and costing less to operate. The model can also process up to 12 times more text at once compared to competing models – a decisive advantage for data-intensive tasks like document analysis or code reviews.
The essentials
- Subquadratic developed SubQ, claiming to have solved a mathematical bottleneck in context scaling
- Claimed advantages: 56x faster, 12x larger context window, performance parity with OpenAI, Google DeepMind, and Anthropic on coding tasks
- So far only self-published test results and independent assessments available; SubQ is not publicly accessible
- Experts compare the situation to Theranos – either revolutionary or massive false claims
The problem: context scaling as AI's bottleneck
Large language models become exponentially more expensive, slower, and energy-hungry as context grows. This is a known Transformer architecture problem that has plagued the industry for roughly a decade. Subquadratic claims to have solved this bottleneck using a technique called Sparse Attention – an approach that processes only the most relevant tokens rather than computing everything.
Proof yes, verification no
The company has begun providing evidence: independent assessments of its technology have been published, and according to t3n, results suggest the claims "may warrant attention." However, SubQ itself remains inaccessible to the public – no one outside Subquadratic can currently test the model.
Skepticism in the AI community runs deep. AI engineer Dan McAteer captured the general reaction:
"SubQ is either the biggest breakthrough since the Transformer … or it's the 'Theranos of AI'."
The comparison references Theranos, the biotech company that raised $900 million from investors by promising revolutionary blood tests – tests that never existed. After 15 years, the fraud was exposed.
What this means for you
If Subquadratic's claims hold up, it would solve one of the biggest bottlenecks in modern AI systems. European enterprises working with large document volumes – insurance companies, law firms, research institutions – could then operate AI systems far more cheaply and quickly. However: without public access and independent lab verification, the promises remain marketing. The coming weeks will reveal whether Subquadratic has a genuine breakthrough or whether the AI community is falling for another case of hype-driven false claims.
Sources
Editorially owned by Ideal Syka. Sources and method: Newsroom & method. Tips and corrections: ai@i6eal.de.
