Subquadratic Announces SubQ Snap, A Decision Model With 384,000-Token Input
The pitch targets the narrow decisions agents make constantly, where Subquadratic says current decision models stall at 32,000 tokens, so a wide-context model at a low per-token price aims at workflow cost and latency rather than at the headline reasoning tasks LLMs already handle.
Reporting from 1 source: GIGAZINE.
Subquadratic announced SubQ Snap, a decision-making model for agent workflows that handles up to 384,000 tokens of text and images, a context window the company describes as twelve times that of Jev. It processes requests of up to 100,000 tokens in under one second, scores four points higher than Jev on Decision Index version 0.2.1, and costs $0.05 per million input tokens with no charge for output.
Subquadratic built SubQ Snap for the small judgments inside agent workflows: which files are relevant, which tool to use next, whether test results pass, and which steps need a state-of-the-art model. The company says existing decision models are stuck at 32,000 tokens, so they judge without seeing most of the material.
SubQ Snap uses SSA, an attention mechanism Subquadratic developed for long-context processing. It handles images as well as text, and the company reports it beats Jev across most of the Decision Index items, with ACOS, API-Bank and When2Call the exceptions. A public preview version is available at the time of writing.
Synthesized by Yomimono from the 1 cited source below, including Japanese-language reporting where cited, then editorially reviewed before publishing.