# Subquadratic's founders have a new answer to the AI efficiency problem: show the benchmarks > Source: > Published: 2026-06-19 11:19:28+00:00 [Justin Dangel](https://www.linkedin.com/in/justin-dangel-b45aa01?ref=runtimewire) and [Alex Whedon](https://www.linkedin.com/in/alexander-whedon?ref=runtimewire) are trying to move Subquadratic's biggest claim out of the realm of launch-day belief and into measured evidence. [Subquadratic](https://subq.ai?ref=runtimewire), a Miami-based AI startup, came out of stealth in May 2026 with [SubQ](https://subq.ai/introducing-subq?ref=runtimewire), a large language model that Dangel and Whedon say replaces the dense attention mechanism at the core of transformer models with a sparse-attention architecture designed to scale more efficiently as context grows. On June 19, [MIT Technology Review](https://www.technologyreview.com/2026/06/19/1139313/a-startup-claims-it-broke-through-a-bottleneck-thats-holding-back-llms/?ref=runtimewire) reported that Subquadratic had followed its thinly evidenced launch with additional disclosures, including third-party tests from Appen. That sequence matters. Subquadratic did not merely launch a long-context product. It challenged the technical foundation of the current LLM market, where models from Google DeepMind, [OpenAI](https://x.com/OpenAI?ref=runtimewire), Anthropic and others still depend on transformer attention. Dangel, Subquadratic's cofounder and CEO, framed the ambition directly to MIT Technology Review: "We hope we're kicking off a new age of efficiency." He added, "We don't think anybody will be building on transformers in a few years." That is the kind of claim founders make when they want investors, customers and researchers to look at the architecture, not just the demo. It is also the kind of claim that invites a hard question: if the model really changes the economics of long-context AI, why was the proof not there on day one? ### The bottleneck Subquadratic is attacking The target is dense attention, the mechanism popularized by the 2017 Google paper ["Attention Is All You Need"](https://arxiv.org/abs/1706.03762?ref=runtimewire). In a standard transformer, each token in a sequence is compared with every other token. That gives the model a powerful way to track relationships across text, but it also means the compute cost rises sharply as inputs get longer. MIT Technology Review uses the example of 10,000 words triggering almost 50 million individual multiplications. Subquadratic's thesis is that not every token-to-token relationship deserves compute. Whedon, Subquadratic's cofounder and CTO, told MIT Technology Review: "If you're reading a book, you're not going to look at the first and second words, first and third - that's insane." Sparse attention tries to select the relationships that matter and skip the rest. The idea is not new. The hard part is preserving model quality while removing most of the attention work. That is why Subquadratic's specific claim is sharper than "longer context." Subquadratic says SubQ is the first LLM built on a sparse-attention architecture that can compete with dense-attention models on important tasks while processing much longer inputs. ### The first launch left a verification gap Subquadratic's May launch paired a large technical claim with limited outside validation. MIT Technology Review reported that the company initially offered little evidence beyond self-published scores, and that SubQ still has not been made widely available for public testing. That left the founders with a credibility problem familiar to any technical startup claiming a step-function improvement: the bigger the delta, the less useful self-reported benchmarks become. Dan McAteer, an AI engineer, captured the early reaction in a post cited by MIT Technology Review: "SubQ is either the biggest breakthrough since the Transformer ... or it's AI Theranos." Whedon acknowledged the mistake in timing. "We expected healthy skepticism," he told MIT Technology Review. "In hindsight, releasing the third-party benchmarks alongside the initial announcement would have preempted much of the skepticism, which is why we're taking the time to make sure any future results are fully verified before putting them out." That is the right founder posture for this market. AI buyers have seen enough benchmark theater to discount charts that are not independently run, methodologically clear and tied to workloads people actually need. Subquadratic's best chance is not to out-market the skepticism. It is to publish enough detail that the technical audience can separate an architecture result from a launch narrative. ### What Appen's tests add The strongest new support is Appen's evaluation. In a June brief, [Appen](https://www.appen.com/whitepapers/subquadratic-preview-model-benchmark-evaluation?ref=runtimewire) said Subquadratic engaged it to run independent tests on SubQ 1.1 Small Preview. Those results, along with additional information the company has released, make the original claim harder to dismiss. Appen's Jeanine Sinanan-Singh told MIT Technology Review that the work "validated their architecture," while also making the point Subquadratic should have led with: shocking results are less credible when a company says them about itself. Subquadratic also published a [technical report for SubQ-1.1-Small](https://subq.ai/docs/subq-1-1-small-model-card.pdf?ref=runtimewire) describing its sparse-attention approach and providing additional benchmarks. Those are company-authored materials, but they are now anchored by a third-party evaluation rather than sitting alone in a launch post. The next proof point is broader access. Until outside researchers and customers can test SubQ across messy repositories, document sets and agent workflows, the claim remains promising but not settled. ### The founder bet is economic, not just technical Subquadratic's bet is ultimately economic as much as it is technical: that the next major AI infrastructure company may not be another wrapper around the same transformer stack, but a model lab that changes the cost curve for context-heavy work. That background explains the product wedge. SubQ is not being pitched first as a chatbot. It is being pitched at problems where context scarcity creates operational friction: reading entire codebases, analyzing document collections, maintaining long-running agent state, and reducing the brittle retrieval and orchestration layers that companies have built around current model limits. If Subquadratic is right, the opportunity is not just a cheaper API call. It is a chance to simplify parts of the modern AI application stack. Retrieval systems, chunking, summarization and multi-agent handoffs exist partly because the model cannot afford to see everything at once. A model that can see more at lower cost changes the build-vs-orchestrate tradeoff for developers. But that is also why the scrutiny will stay high. Sparse attention has been tried before. Dense attention remains dominant because quality losses have usually erased the efficiency gains. Subquadratic's founders are asking the market to believe they have found a version that keeps the upside and removes the penalty. The Appen tests make that claim credible enough to investigate. They do not make it inevitable. Subquadratic has moved from assertion to evidence. The next phase is exposure: put SubQ in enough independent hands that the model is judged by real workloads, not launch math.