Subquadratic Launches SubQ Snap Decision Model With 384K Token Context

SubQ Snap Joins the Decision Model Race

San Francisco-based startup Subquadratic has launched SubQ Snap, a decision model it says can process up to 384,000 tokens of text and images in a single request, 12 times the 32K-token limit of Jev, the decision model unveiled by TypeSafe in mid September. OpenAI, Perplexity, Cloudflare, Amazon and Microsoft have all launched alternatives since.

Unlike traditional large language models that generate text, decision models select from predefined answers, such as which tool an agent should call or whether a test passed. Subquadratic says Snap can process 100,000 tokens and return a decision in about one second, potentially letting agents work across large codebases and documents without splitting information into chunks.

Sparse Attention and Benchmarks

Snap is built on Subquadratic Sparse Attention (SSA), introduced in May, which the company says grows attention requirements linearly rather than quadratically with context length. CEO Alex Whedon told The New Stack that Snap was developed with inspiration from Jev and that decision models remove the need for decoding for some problems.

Subquadratic reports Snap scored 96.3 to 99.8 on an adapted RULER benchmark from 32K to one million tokens, and 74.0 on the Decision LCR test against Jev's 48.0. On Decision Index 0.2.1, Snap scored 61.7 against 57.9, and 82.9% on Bespoke Labs' Nimble benchmark against 76.0%. Across 16 enterprise benchmarks, Snap led on 13, averaging 67.6 versus 61.6, though Jev won three tests including When2Call.

Preview Limits and Pricing

Whedon confirmed all results were produced internally, though the code for adapted long-context evaluations was released. The 384K limit is a temporary restriction for the public preview; Snap was tested at one million tokens but supporting that publicly would require about 2.6 times more compute.

Snap is available in public preview at $0.05 per million input tokens with no charge for outputs, making a 100K request about half a cent and a full 384K request about two cents.

Company Background

Subquadratic emerged from stealth in May with $29 million in seed funding at a reported $500 million valuation, previously operating as Aldea. Its debut model SubQ claimed a 12-million-token context window. Snap extends that long-context research into the crowded decision model market.

Sources: The New Stack