“Take work away from frontier LLMs”: Subquadratic’s decision model claims 12x Jev’s context window
The New Stack Paul Sawers
Subquadratic says its new decision model can read 384K tokens at once. That’s 12x Jev’s window, and it’s aimed at taking work off big LLMs.
Based on reporting by The New Stack, Paul Sawers — read the original for the full story.
Summary, retelling and take written by AI under human oversight; images are AI-generated illustrations. How we work · Report an error
Subquadratic has jumped into the decision-model pileup with a big claim: its new model, Snap, can handle up to 384,000 tokens of text and images in one request. That’s 12 times Jev’s 32K-token decision window, and the company says the model can process 100K tokens and return a decision in about a second.
The pitch is simple. Decision models don’t write essays or code; they pick from predefined answers. That might mean choosing the next tool for an agent, finding the right file, or deciding whether a test passed. Subquadratic’s CEO and co-founder, Alex Whedon, says the point is not just to support frontier LLMs, but to shift work away from them.
His argument is that current decision models hit a wall when the context gets too large. A coding agent chasing a bug may need code, logs, tests, and docs at once, but a 32K window can force painful trimming. Snap’s larger window, Whedon says, lets it inspect more of the relevant material before making a choice. The company’s sparse-attention setup, called SSA, is meant to keep that long-context processing from becoming absurdly expensive.
Subquadratic is backing the story with its own benchmarks. On an adapted RULER retrieval test, Snap scored between 96.3 and 99.8 across inputs from 32K to one million tokens. On Decision LCR, it scored 74.0 versus Jev’s 48.0, and on 16 enterprise benchmarks it says Snap beat Jev on 13, with an average score of 67.6 to 61.6. Jev still won three tests, including one called When2Call.
There’s a catch, of course. Whedon says all the results were produced internally, although the company has released code for some of its long-context evaluations. Snap is in public preview, not on the latest Decision Index leaderboard, and its 384K limit is temporary because of compute constraints. Subquadratic says it tested the model at one million tokens, but offering that publicly would take about 2.6 times more compute.
My take — AI-written commentary, not fact-checked reporting
This is the right fight. If a decision model can chew through whole codebases instead of little chopped-up scraps, that’s not hype, that’s less plumbing. The funny part is how quickly the industry reinvented “read more, decide better” and called it a new category.
Read more about this at: The New Stack