Subquadratic releases SubQ Snap, a decision model with a 384,000 token window

An agent that picks which tool to call next does better when it has read everything that matters. Most decision models today take about 32,000 tokens of context, so an agent working across a repository or a long contract has to cut the evidence down first, and the step that chooses what to keep is where mistakes happen. Subquadratic released SubQ Snap on 12 October, a decision model that it says takes up to 384,000 tokens of text and images in a single call, twelve times the Jev context limit from TypeSafe.

What Subquadratic says the model does

Subquadratic says Snap returns a decision on a 100,000 token request in about one second. The company lists four roles inside an agent loop: routing a request to the right path, picking the next tool, checking whether a result passes, and deciding when to stop. The context accepts text and images in the same request, so a screenshot or a scanned page can be read next to the surrounding documents and logs.

The architecture is Subquadratic Sparse Attention, which the company first described in May. It replaces a standard transformer's dense self-attention, whose compute cost grows with the square of context length, with content-dependent selection the company says grows linearly with context length. A decision head then maps the representations to a probability over the allowed answers, so Snap does not generate text.

The numbers Subquadratic publishes

Every figure is Subquadratic's own, run on its own harness. On an adapted long-context retrieval test called Decision RULER, Snap scores between 96.3 and 99.8 across inputs from 32,000 to 1 million tokens, where 100 is perfect. On a multi-document reasoning test called Decision AA-LCR, Snap scores 74.0 against Jev's 48.0, with Jev processing the documents in chunks because its context limit does not hold them. On Decision Index 0.2.1, where every request fits inside 34,000 tokens, Snap scores 61.7 against Jev's 57.9.

Across 16 Decision Index benchmarks the company picked as enterprise workflows, Snap averages 67.6 against Jev's 61.6 and leads on 13 of the 16. The gap is largest on phishing detection (47.1 against 25.1), code output prediction (73.8 against 57.1) and legal clause review (70.7 against 59.1). Jev leads on three tests, including When2Call, which checks whether an agent should call a tool, ask for help, answer or decline (67.0 against 74.6). On Bespoke Labs' Nimble suite of 3,880 labelled items, Snap scores 82.9 percent against Jev's 76.0 percent.

What stays unverified

The 384,000 token limit is a temporary cap for the preview. Subquadratic says it tested Snap at one million tokens and that supporting that length publicly would need about 2.6 times more compute. Snap is not on the public Decision Index leaderboard: the maintainers told the company that the current 0.3.2 release is for open-weight models with Jev as an exception, and that private tests in it mean the 0.2.1 score Subquadratic reports is not directly comparable. Snap costs $0.05 per million input tokens, with output tokens free. The company has released the code for the long-context evaluations it adapted so that others can reproduce the numbers.

A decision model is a trade: the model does not write a sentence, so the application lists every answer first. A longer context widens the set of questions where a decision model can replace a frontier call, because the evidence no longer has to be summarised first. The next step for a team already running a decision model is a side-by-side run on the same prompt, with the chunking step removed and the full context handed to Snap, measuring decision quality and cost against the current setup.

Source

Introducing SubQ Snap: A Long-Context Decision Model for Enterprise Agents on Subquadratic's blog. The New Stack covered it in Take work away from frontier LLMs: Subquadratic's decision model claims 12x Jev's context window, with additional comment from Subquadratic CEO Alex Whedon.