We build frontier models
that scale efficiently.

Subquadratic is a frontier AI research company building the most compute-,
memory- and sample-efficient algorithms and models.

Built by researchers and engineers from
  • Meta
  • Google
  • Oxford
  • Cambridge
  • MIT
  • BYU
  • DoorDash
  • Amazon
  • LlamaIndex
  • TwelveLabs
  • Airbnb
Language Model – Early Access

SubQ Savant

The first long-context language model for data-intensive work. Savant reasons across millions of tokens in a single request, at a fraction of the cost.

Learn More

Coding across a repository

Agents work with the whole repository on changes that span files and dependencies, such as a legacy migration.

Long-running agents

Agents carry the full task history and earlier tool results from one step to the next.

Document analysis

Teams retrieve evidence, connect facts across documents and trace each answer back to its source passages.

Decision Model – Available Now

SubQ Snap

The first long-context decision model for real enterprise work. Snap handles the context-heavy, structured decisions in your agentic workflows: routing, ranking, classifying and tool calls, in seconds, at $0.04 per 1M tokens.

Learn More

Decisions with long, multimodal context

Snap reads up to 384K tokens, enough for the whole repo, data room or agent history, so every call is grounded in all the context it needs.

Routine decisions at $0.04 per 1M tokens

Most agent steps are simple calls: pick a tool, open a file, check a result. Snap handles them and leaves only the hard work to your LLM.

Decisions in seconds, not minutes

Under a second for 100K tokens, so your agents move faster without giving up a single point of accuracy.

Rewriting the Scaling Laws of AI.

Today's transformer-based LLMs waste compute by processing every possible relationship between tokens, but only a small fraction of these relationships matter. This is called quadratic scaling.

SubQ is built differently. Using a proprietary algorithm called Subquadratic Sparse Attention (SSA), it isolates the tokens and relationships that matter, ensuring compute is used efficiently.

Read the Research
0 128K 512K 1M 2M TOKENS 8× 31.5× 64.5× 128× 1,008 PFLOP 7.8 PFLOP Frontier Models (Dense Attention) SubQ (SSA)
Enterprise

Interested in
enterprise access?
Contact sales.

We'll use your details to reply to your request. See our privacy policy.
Research

From the lab.

SubQ Snap 1.2: Order-Robust Decisions and Higher Long-Context Accuracy
Product · November 11, 2026
SubQ Snap 1.2: Order-Robust Decisions and Higher Long-Context Accuracy
Read more →
Introducing SubQ Snap 1.1: The Smartest, Enterprise-Ready Decision Model
Product · October 22, 2026
Introducing SubQ Snap 1.1: The Smartest, Enterprise-Ready Decision Model
Read more →
Introducing SubQ Snap: A Long-Context Decision Model for Enterprise Agents
Product · October 12, 2026
Introducing SubQ Snap: A Long-Context Decision Model for Enterprise Agents
Read more →
Introducing SubQ 1.1 Small
Research · June 16, 2026
Introducing SubQ 1.1 Small
Read more →

Building a radically more intelligent world

Subquadratic is a frontier AI research company that believes the architecture layer is the highest-leverage surface in AI. While other major labs focus on incremental improvements to Transformer models, we're pushing foundational change at the architecture level to build models and products that scale efficiently.

Built by researchers and engineers from
  • Meta
  • Google
  • Oxford
  • Cambridge
  • MIT
  • BYU
  • DoorDash
  • Amazon
  • LlamaIndex
  • TwelveLabs
  • Airbnb