Skip to content

Model releases

SubQ Snap reads 384K tokens before each agent decision

7 min read

Written by an AI model from the linked sources and checked against them by a second model. How we work

Image: zeniteq.com

Subquadratic has entered the decision-model race with SubQ Snap, a model that picks an agent’s next move after reading up to 384,000 tokens of text and images. That is twelve times the 32K that TypeSafe’s Jev takes per decision, the company said in an October 12 post on the Subquadratic site.

Subquadratic says Snap answers a 100K-token request in about a second. It costs $0.05 per million input tokens, and output is free. The model is in public preview. The company describes the 384K ceiling as a temporary preview limit set by the compute it has available. Snap also reads screenshots, charts, diagrams and scanned pages in the same context as the documents and logs around them.

Where a decision’s cost sits

A decision model gets some context, a question and a list of allowed answers fixed in advance. It returns a probability for each answer and writes no text. Subquadratic calls this narrowing of the output “specialization”, and the post presents it as an interface rather than an architecture. Jev, Perplexity’s Decider and Laya all offer the same contract: a question goes in and a structured prediction over a finite set comes out. They are built differently, though. Laya pairs a bidirectional encoder with a decision head. Decider v1.1 takes a pre-trained language model and switches it to non-causal attention, which lets every token see every other token.

Subquadratic’s argument starts there. Dropping generation removes the cost of the output, but the model still has to read the input. For a decision model nearly all inference compute is prefill, the single pass over the prompt before any answer comes out. Under dense self-attention every token is compared with every other, so doubling the context quadruples attention compute. That is why, the post says, today’s decision models stay at 32K tokens “and decide without most of what they need to see.”

The usual workarounds are retrieval, summarization, truncation and chunking. Each one decides upstream what the model will see. The post’s examples of what gets lost are specific: a user constraint stated 40 turns earlier, a function defined in another module of a large codebase, a qualifying clause in a separate document. Every extra step in that pipeline also adds latency.

Snap handles the two costs separately. Subquadratic Sparse Attention (SSA) deals with reading. It skips most of the token-pair comparisons and keeps the ones that matter for a given input, and the company says this turns quadratic scaling into roughly linear scaling. Which comparisons it keeps depends on the content, where older sparse-attention work used fixed positional patterns. Subquadratic’s Alexander Whedon has put it this way: “For prompt A, words one and six are going to be important to each other. For prompt B, maybe it’s words two and three. It’s different for every prompt.” A decision head then produces the answer. It maps the model’s internal representations directly to a probability distribution over the admissible answers, with no token-by-token decoding.

SSA already powers SubQ, which Subquadratic calls the first fully subquadratic LLM. SubQ launched earlier this year with $29 million in seed funding and a 12 million token context window. Snap applies the same research to a much smaller job. The company pitches it at four recurring calls inside an agent loop: Routing, Tool Picking, Checking and Stopping.

Four points over Jev, three losses

Every score below comes from Subquadratic’s own runs. On the public suite of Decision Index v0.2.1, where every request is under 34K tokens, Snap scores 61.7 to Jev’s 57.9. It leads in four of five subareas: arts and human taste by 12.11 points, language understanding by 7.46, retrieval and classification by 6.27, and tools and automation by 3.39. On Bespoke Labs’ Nimble Suite 13, a macro average over 13 decision sets, Snap scores 82.9% against Jev’s 76.0%.

Subquadratic also picked out 16 Decision Index tasks that it says map most closely to enterprise work. Snap averages 67.54 on them (the post’s summary rounds this to 67.6) against Jev’s 61.6, and it leads on 13 of the 16. The widest gaps are in phishing detection on PhishNChips (47.1 against 25.1), predicting code output on CRUXEval (73.8 against 57.1), RAG hallucination detection on RAGTruth (64.8 against 51.3), entity sentiment in financial news on FinEntity (92.8 against 80.8) and full-NDA clause review on ContractNLI (70.7 against 59.1).

Jev wins three. On When2Call, which tests whether a model should call a tool, ask, answer or decline, Jev scores 74.6 to Snap’s 67.0. It also edges ahead on API-Bank’s next-API-call task (88.0 to 86.4) and on ACOS aspect-level review analysis (27.3 to 26.7). Two of those three losses are in the tool-use category.

The long-context tests are where Snap is meant to pull away, and Jev mostly can’t enter them. Decision RULER covers 11 kinds of long-context retrieval task. Snap scores 99.8 at 32K, where Jev scores 94.1. It then scores 99.8 at 64K, 99.5 at 128K and 99.8 at 256K, and Jev has no entry beyond 32K. Decision LCR tests reasoning that combines evidence from different parts of an artifact, on inputs of 72K to 115K tokens. Snap reads the whole input in one shot and scores 74.0. Jev, run with chunking, scores 48.0. That comparison sets Snap against the workaround developers already use with Jev today.

Price per token, value per call

Per token, Snap costs slightly more than Jev, which launched in September at $0.042 per million input tokens. So Snap’s case on price rests on how much fits into one call. A bounded answer space does not mean a small input, the post argues: a choice among just three tools can still depend on evidence spread across a codebase or a long interaction history. Chunking that evidence adds steps and latency, and on Subquadratic’s LCR numbers it also produces worse answers.

For a coding agent, Subquadratic suggests sending tool selection and result verification to the decision model. The larger generative model stays on writing code and on extended reasoning.

“A vacuum of vision”

The field Snap joins is barely a month old. According to The New Stack, TypeSafe unveiled Jev in mid-September, and OpenAI, Perplexity, Cloudflare, Amazon and, most recently, Microsoft have all joined since. The outlet also reported on October 13 that Workday uses Jev to build interfaces on the fly for its Sana AI assistant. Jev picks prebuilt components such as paycheck cards and date pickers based on what an employee asks.

In an email Q&A with The New Stack, Whedon, whom the outlet identifies as Subquadratic’s CEO and co-founder, said Jev shaped the company’s entry. “Snap was developed with inspiration from Jev, because decision models are highly thesis-aligned,” he said. “Decision models remove the need for decoding for some problems.”

His complaint is with how the category gets used. Most decision models today work around the LLM, the post says, judging its outputs or choosing which LLM handles a task. Subquadratic wants Snap to take on decisions inside workflows that LLMs make today, and “even in some cases complete large-scale tasks on their own.” Whedon put it more bluntly.

“We see a vacuum of vision for decision models. Decision models become exciting when they actually take work away from frontier LLMs.”

No comments yet. Be the first.

Join the conversation

Sign in to leave a comment. It appears after moderation.