The first model built for long‑context tasks
The Hook
“Today we're announcing a major architectural breakthrough. The world's first fully subquadratic LLM. SubQ has a 12 million token context window, and frontier-level intelligence.”
Product Reveal
“It outperforms Opus 4.7 on long context at less than 5% of the cost.”
Feature Teaser
“At 1 million tokens, SubQ processes tokens 52 times faster than Flash Attention. It dramatically changes the cost of training and inference, enabling us to scale context windows to 12 million tokens and beyond, pushing the boundaries of intelligence.”
Feature Teaser
“This means your agents can now complete weeks of work at a time without degrading. Reasoning across entire codebases, merging hundreds of PRs at once, and finding patterns in tens of thousands of documents. All in one shot, without losing accuracy, speed, or context.”
Product Reveal
“SubQ is not just another model, it represents a major algorithmic breakthrough. It is the first model built on a fully sub-quadratic, sparse-attention architecture.”
Problem Agitation
“LLMs today waste compute by processing every possible relationship between words. But only a small fraction of these matter. SubQ finds and focuses only on those relationships, meaning compute is used where it matters most. At 12 million tokens, this reduces compute almost 1,000 times, changing the way that LLMs scale.”
Call to Action
“Today, SubQ is available for early access, along with SubQ Code, a coding agent built on the next generation of AI technology. This marks a new era for builders. An era powered by subquadratic intelligence. Visit Subq.ai to get early access.”
Synthesizable in Remotion · key components of the subquadratic-launch-explainer template.
Circular node cluster transitioning from full radial edges to sparse selective links.
Generate node positions along an SVG circle. Render connecting lines whose opacity interpolates based on Remotion frame state.
Multi-colored square blocks populating the outer screen perimeter.
Build a flex grid border populated with animated color cells using spring-driven opacity and scale.
Side-by-side vertical bar graphs rising to distinct target heights with top label tags.
Animate SVG rect height using interpolate() with spring damping, rendering percentage tags above the final heights.
Purpose. Hook technical audiences with state-of-the-art context window specs.
Execution. Presenter speaking framed by colorful pixel-grid edges and large context numbers.
Purpose. Validate model strength using standard industry benchmarks.
Execution. Dark-mode bar graphs showing SWE-bench and long-context retrieval scores.
Purpose. Explain how sparse attention enables 1000x compute reduction.
Execution. Interactive node graph shifting from wasteful dense connections to pruned sparse paths.
Purpose. Direct users to register for early access.
Execution. Clean serif text URL over presenter resolving to the pixel logo on black.