Subquadratic launches SubQ Snap, a long context decision model it says keeps compute per token constant as context grows
Introducing SubQ Snap, the first long-context decision model for real enterprise work. Snap is the only decision model or LLM where compute per token stays constant as context grows. • 384K tokens of multimodal context, 12x larger than Jev. • 61.7 on Decision Index v0.2.1 vs Jev’s 57.9, outperforming on 13 of 16 enterprise benchmarks. •Processes 100K tokens and returns a decision in under one second. •$0.05/MM input tokens. Output is free. Available today to anyone. The first 10,000 sign-ups get 100M free input tokens. We’ll likely hit capacity fast and will add users as we add compute. Compute offers welcome 🙂. Latency will be variable as we tune the platform to this new model and category of users. Model speeds are 80ms for minimal context, 250ms for 32k, 660ms for 100k, and 2.5 seconds for 384k. Visit:
Context
Decision models handle the fast routing and screening calls inside AI agents, and a much longer context window at low cost could let them judge whole documents instead of snippets.
Subquadratic is an AI research company that emerged from stealth this year with SubQ, which it called the first LLM built on a fully subquadratic architecture, where compute grows linearly with context length. Decision models return a choice, score or label with probabilities instead of open ended text. Jev is the decision model that SubQ compares itself against throughout its announcement.
Deep dive
On October 12, Alexander Whedon introduced SubQ Snap, calling it the first long context decision model for real enterprise work and the only decision model or LLM where compute per token stays constant as context grows.
He listed 384K tokens of multimodal context, 12 times larger than Jev; a score of 61.7 on Decision Index v0.2.1 against Jev's 57.9, outperforming on 13 of 16 enterprise benchmarks; decisions on 100K tokens in under one second; and pricing of $0.05 per million input tokens with free output. The first 10,000 sign ups get 100M free input tokens, and he warned latency will vary, giving speeds of 80ms at minimal context up to 2.5 seconds at 384K.
According to Subquadratic's press release, Snap runs on Subquadratic Sparse Attention and scores 96.3 on Decision RULER at 1M tokens and 74.0 on Decision LCR, against 48.0 for Jev with chunking. The release said Snap averages 67.6 to Jev's 61.6 across 16 Decision Index benchmarks, while the company's site lists 67.5.
The release quotes Whedon, identified as cofounder and CEO, saying today's decision models have context windows too small or slow for real enterprise work. All benchmark figures come from the company.
What to expect next
Watch for independent tests of Snap's long context accuracy and latency, especially whether constant compute per token holds up under real enterprise loads, and how fast Subquadratic adds capacity after warning it would hit limits.
Also watch how Jev's makers and other decision model teams, such as vLLM Semantic Router with Decision 3.0, respond on context length and price, and whether enterprise buyers adopt it for routing and screening work.