With Snap, Subquadratic aims to reduce the cost of AI decision-making.

Subquadratic reveals SubQ Snap, a multimodal decision model capable of processing 100,000 tokens in under a second, with a 384,000-token context.

Choosing a tool, verifying a document, or deciding on the next step of a task: these operations do not always require calling upon a large language model. Based on this observation, Subquadratic developed SubQ Snap, unveiled on October 12. This specialized decision-making AI model promises to process large volumes of information without experiencing the slowdowns usually associated with long contexts. Fast decisions, even with a large context

SubQ Snap accepts up to 384,000 tokens of text and images in a single query. This is enough to examine a vast set of documents, computer code, contracts, or the action history of an AI agent before selecting a response from several possibilities.

According to metrics shared by Subquadratic, the model can process 100,000 tokens and render a decision in under a second, with a reported latency of 660 milliseconds. At full capacity, or 384,000 tokens, this delay increases to about 2.5 seconds. The company attributes these performances to its Subquadratic Sparse Attention (SSA) architecture, designed to prevent computational requirements from growing quadratically with context length. Superior results to Jev across several benchmarks

On the Decision Index v0.2.1, SubQ Snap achieves a score of 61.7 compared to 57.9 for Jev, another model specialized in structured decisions. Subquadratic also announces a lead on 13 of the 16 tests dedicated to professional use cases, notably in phishing detection, contract analysis, document verification, and code.

On a multi-document reasoning evaluation totaling between 72,000 and 115,000 tokens, Snap reaches 74 points compared to 48 for Jev, the latter utilizing document chunking. These results come from tests published by Subquadratic and remain to be confirmed under independent usage conditions. A model designed to work alongside large LLMs

The goal is not to replace generalist models, but to save them from certain repetitive operations. In a software development agent, Snap could determine which file to check, which tool to run, or whether a result is satisfactory, while reserving generation tasks and complex reasoning for more powerful models.

SubQ Snap is available in public preview, with a announced price of $0.05 per million input tokens and free outputs. The first 10,000 registrants can receive 100 million free input tokens. Access remains subject to available capacity, with a potential waitlist.

With Snap, Subquadratic joins the emergence of specialized decision models. Their value will notably depend on their ability to maintain high accuracy over extended contexts without causing processing costs to explode. A particularly concrete issue for AI agents required to make numerous decisions over the course of a single task.