Subquadratic has entered the decision-model race by launching SubQ Snap in public preview, saying that the model can process up to 384,000 tokens of text and images in a single request. That is 12 times the 32,000-token context limit of TypeSafe’s Jev model, according to the company’s comparison.
Subquadratic says Snap can process 100,000 tokens and issue a decision in about one second, enabling AI agents to inspect large portions of code repositories, documentation, and interaction logs before determining the next action.
What Do Decision Models Do?
Unlike large language models, which generate text and code, decision models typically select an answer from a predefined set. This may include choosing the next tool for an intelligent agent, identifying files related to an issue, or determining whether a test has passed.
Alex Whedon, CEO and co-founder of Subquadratic, believes these models should not be limited to evaluating language-model outputs or selecting the most suitable model for a task. Instead, they should handle work that can eliminate the need to use more expensive large language models.
Longer Context and Internal Tests
Snap is based on an architecture the company calls Subquadratic Sparse Attention, which it says makes attention requirements grow linearly with context length rather than quadratically, as is common in transformers. The company says the model was tested up to one million tokens, but the public preview is currently limited to 384,000 tokens because of available computing capacity; supporting one million tokens requires about 2.6 times the computing resources, according to Whedon.
In a modified version of the RULER information-retrieval test, Snap scored between 96.3 and 99.8 with inputs ranging from 32,000 to one million tokens. In the Decision LCR test for reasoning across multiple documents, it scored 74.0 versus 48.0 for Jev, which processed the documents in batches. Snap also scored 61.7 versus 57.9 on Decision Index 0.2.1, and achieved 82.9% versus Jev’s 76.0% on Bespoke Labs’ Nimble test.
Across 16 enterprise-specific tests, Subquadratic says Snap outperformed in 13 tests, with an average score of 67.6 versus Jev’s 61.6. However, it lost in three tests, including When2Call, which measures an agent’s decision between calling a tool, requesting clarification, answering, or refusing.
What Changes in Practice?
The larger context may reduce the need to chunk documents, summarize them, and perform precise retrieval when the data falls within the 384,000-token limit. Whedon says this allows complete documents or interconnected software modules to be entered instead of selecting small snippets that may leave out important information.
However, the published results are not an independent evaluation; all of them were produced internally by Subquadratic, although the company has made the long-context test code available for reproduction. Snap is also absent from the latest Decision Index leaderboard, because the current release focuses on open-weight models and includes private tests, preventing a direct comparison between Snap’s result in version 0.2.1 and the newer results.
Availability and Cost
Snap is currently available in public preview at $0.05 per million input tokens, with no charge for outputs. The company estimates that processing 100,000 tokens costs about half a cent, while using the full 384,000-token window costs about two cents.
Subquadratic explains that retrieval will not disappear, because enterprise datasets are larger than the available context capacity. However, the company is betting that a larger window will allow complete documents or modules to be retrieved instead of relying on precise and fragile snippet selection.