Subquadratic, a foundational model company focused on improving AI scaling laws, announced today the launch of SubQ Snap, a multimodal decision model built for fast, long-context decisions in enterprise agentic workflows. Powered by Subquadratic Sparse Attention, it has constant compute cost per token, while models like Jev or Claude have quadratically exploding compute costs and inverse economics of scale. This enables it to reason over substantially more of an artifact without giving up the speed and efficiency that make decision models useful.

Highlights include:

  • Up to 384K tokens of context, 12x larger than Jev. Snap can reason over long codebases, documents, interaction histories, and other artifacts in a single call, without requiring truncation or chunking.

  • Fast at long context. Snap can process 100K tokens and return a decision in ~1 second.

  • Multimodal. Snap accepts text and images in the same context, allowing screenshots, charts, diagrams, and scanned pages to be reasoned over alongside the surrounding documents and logs.

  • Category leading performance. Snap scores 96.3 on Decision RULER at 1MM tokens, and 74.0 on Decision LCR, compared to Jev’s 48.0% aided by chunking. Across 16 Decision Index benchmarks spanning tool use, security, legal, finance, support routing, and code, Snap averages 67.6 versus Jev’s 61.6 and leads on 13 of 16 tasks.

  • Achieves frontier performance on standard short-context evals. SubQ Snap scores 61.7 on the Decision Index public suite, where all requests are under 34k tokens, while Jev scores 57.9. It leads Jev in four of the five subareas, including arts & human taste (+12.11), language understanding (+7.46), retrieval & classification (+6.27), and tools & automation (+3.39).

“Decision models get exciting when they can take real work away from LLMs, but the context window size of today’s decision models are simply too small or slow to handle real enterprise work,” said Alex Whedon, co-founder and CEO of Subquadratic. “That’s why we created SubQ Snap. To create a radically more intelligent world we will need radically different ways to scale AI, not just more capital, and Snap is just another step in a more scalable direction.”

Taking real work off generative models

Unlike traditional large language models designed for open-ended generation, decision models are optimized to choose from a predefined set of options, such as which tool to use next, whether a result passes, or which step should be handed to a larger model. Agents make these decisions repeatedly throughout a workflow, so moving them to a fast, specialized model can reduce the cost and latency of the system as a whole.

But decision models only take meaningful work off larger models if they have enough context to make those decisions correctly. In enterprise workflows, the relevant evidence may be spread across a repository, a long interaction history, a contract and its exhibits, or several related documents. If the decision model cannot process that context, the input has to be reduced, retrieved, or handed back to a larger model.

Snap was built to move that boundary. Its 384K context window and SSA architecture allow it to make bounded decisions over substantially larger artifacts while maintaining the speed and efficiency expected from a decision model. In Subquadratic’s evaluations, Snap leads Jev on 13 of 16 enterprise-focused Decision Index benchmarks and maintains strong performance at long context, including a score of 96.3 on Decision RULER at 1mm tokens and 74.0 on Decision LCR over real multi-document inputs.

The result is a broader role for decision models inside enterprise agents. More routing, verification, tool selection, and control decisions can be handled directly by a specialized model, while generative models can be reserved for synthesis, coding, and deeper reasoning.

SubQ Snap is part of the SubQ model family which Subquadratic introduced after emerging from stealth earlier this year. SubQ was the first LLM built on a fully subquadratic architecture, where compute grows linearly with context length instead of quadratically. In research testing at 12 million tokens, the architecture cut attention compute by almost 1,000x compared with other frontier models.

Learn more about SubQ Snap here.

SubQ Snap is available today for limited access, please visit Subq.ai for access.

About Subquadratic

Subquadratic is a frontier AI research company building the most compute-, memory- and sample-efficient algorithms and models. Our mission is to build a radically more intelligent world to augment human potential. This requires rewriting the scaling laws of AI, not just more capital.

Media gallery

About The Author
Business Wire