A little-known Miami-based startup named Subquadratic has emerged from stealth mode with a bold claim: it has built the first large language model (LLM) that completely sidesteps the mathematical constraint that has defined—and limited—every major AI system since 2017. The company asserts that its new model, SubQ 1M-Preview, is the first LLM built on a fully subquadratic architecture, where computational demand grows linearly with context length rather than quadratically. If validated, this would mark a genuine inflection point in AI scaling, potentially reducing compute requirements by nearly 1,000 times for long-context tasks. While the AI community is intrigued, many researchers are demanding independent proof before celebrating.

The Quadratic Bottleneck That Defined AI

To understand why Subquadratic's claim is so significant, one must first grasp the fundamental inefficiency at the heart of every modern transformer-based model. Since the seminal 2017 paper "Attention Is All You Need," virtually all frontier AI systems—from OpenAI's GPT-4 to Anthropic's Claude and Google's Gemini—have relied on an operation called attention. Attention compares every token (word or subword) in an input against every other token. While this enables models to capture complex relationships, it comes with a steep cost: as the input length grows, the number of comparisons—and the compute required—increases quadratically.

What Is the Attention Mechanism?

In simple terms, attention allows a model to weigh the importance of each token relative to all others. For a 1,000-token input, that means roughly 1,000² = 1,000,000 interactions. For 100,000 tokens, the number skyrockets to 10 billion. This scaling relationship, known as O(n²) complexity, has shaped the economics of the AI industry. It dictates how long a context window a model can handle and at what cost.

The Cost of Scaling

The industry standard has settled on context windows of 128,000 tokens for many models, with some frontier systems like Claude Sonnet 4.7 and Gemini 3.1 Pro pushing to 1 million tokens. But even at those sizes, processing long inputs becomes punishingly expensive. To cope, developers have built an elaborate stack of workarounds: retrieval-augmented generation (RAG) systems that pull only a few relevant chunks, prompt engineering techniques, multi-agent orchestration, and more. All these approaches exist to route around the fundamental constraint that the model cannot efficiently process everything at once. Subquadratic argues that these workarounds are expensive, brittle, and ultimately limiting.

Subquadratic's Solution: A Linear Scaling Architecture

Subquadratic claims to have broken free of the quadratic chains with its SubQ model. The company says its architecture achieves linear scaling, meaning that doubling the input length only doubles the compute—not quadruples it. If true, this would be a transformational shift. For a 12-million-token input—a context window far beyond anything currently available—Subquadratic says its approach reduces attention compute by almost 1,000 times compared to other frontier models.

Claims of 1,000x Reduction

The numbers the company is publishing are extraordinary. At 12 million tokens, SubQ 1M-Preview would use about 0.1% of the attention compute of a standard transformer handling the same input. This efficiency gain dwarfs any existing approach, from sparse attention to state-space models. However, such a dramatic improvement has raised eyebrows. The AI research community's reaction has been mixed, ranging from genuine curiosity to open accusations of vaporware.

Products and Funding

Alongside the model announcement, Subquadratic launched three products into private beta:

• API exposing the full 12-million-token context window
• SubQ Code, a command-line coding agent
• SubQ Search, a search tool leveraging the long-context capabilities

$29 Million Seed Round

The startup has raised $29 million in seed funding from notable investors, including Tinder co-founder Justin Mateen, former SoftBank Vision Fund partner Javier Villamizar, and early investors in Anthropic, OpenAI, Stripe, and Brex. According to The New Stack, the round values Subquadratic at $500 million—a huge valuation for a company that only just emerged from stealth.

Skepticism and the Need for Independent Verification

While the claims are impressive, the lack of publicly available independent benchmarks has fueled skepticism. The company has not released detailed technical papers or open-sourced the model for rigorous testing. Researchers point out that numerous attempts to achieve subquadratic attention—such as Linformer, Reformer, and Performer—have either failed to scale or compromised on quality. Subquadratic's advance would need to demonstrate both efficiency and accuracy on standard tasks to be taken seriously.

Research Community Reaction

Many leading AI scientists are waiting for peer-reviewed validation. Some have noted that the company's description of its architecture is vague, and the promised efficiency gains seem almost too good to be true. Without transparent benchmarks, the claims remain just that—claims. Subquadratic has said it plans to release more details soon, but for now, the burden of proof lies with them.

The potential is enormous: if Subquadratic's model holds up to scrutiny, it could render much of the current AI infrastructure obsolete, enabling models to process entire books, codebases, or datasets in a single pass. Until then, the community watches with cautious optimism, hoping for a breakthrough—and demanding the data to back it up.