A bit of-known Miami-based startup referred to as Subquadratic emerged from stealth on Tuesday with a sweeping declare: that it has constructed the primary giant language mannequin to completely escape the mathematical constraint that has outlined — and restricted — each main AI system since 2017.
The corporate claims its first mannequin, SubQ 1M-Preview, is the primary LLM constructed on a completely subquadratic structure — one the place compute grows linearly with context size. If that declare holds, it could be a real inflection level in how AI programs scale. At 12 million tokens, the corporate says, its structure reduces consideration compute by nearly 1,000 instances in comparison with different frontier fashions — a determine that, if validated independently, would dwarf the effectivity positive aspects of any present method.
The corporate can also be launching three merchandise into non-public beta: an API exposing the total context window, a command-line coding agent referred to as SubQ Code, and a search instrument referred to as SubQ Search. It has raised $29 million in seed funding from buyers together with Tinder co-founder Justin Mateen, former SoftBank Imaginative and prescient Fund accomplice Javier Villamizar, and early buyers in Anthropic, OpenAI, Stripe, and Brex. The New Stack reported that the elevate values the corporate at $500 million.
The numbers Subquadratic is publishing are extraordinary. The response from the AI analysis neighborhood has been, to place it mildly, blended — starting from real curiosity to open accusations of vaporware. Understanding why requires understanding what the corporate claims to have solved, and why so many prior makes an attempt to resolve the identical downside have fallen brief.
The quadratic scaling downside has formed the economics of the whole AI business
Each transformer-based AI mannequin — which incorporates nearly each frontier system from OpenAI, Anthropic, Google, and others — depends on an operation referred to as "consideration." Each token is in contrast towards each different token, in order inputs develop, the variety of interactions — and the compute required to course of them — scales quadratically. In plain phrases: double the enter measurement, and the associated fee doesn't double. It quadruples.
This relationship has formed what will get constructed and what doesn't. The business commonplace is 128,000 tokens for a lot of AI fashions and as much as 1 million tokens for frontier cloud fashions resembling Claude Sonnet 4.7 and Gemini 3.1 Professional.
Even at these sizes, the price of processing lengthy inputs turns into punishing. The business constructed an elaborate stack of workarounds to manage. RAG programs use a search engine to tug a small variety of related outcomes earlier than sending them to the mannequin, as a result of sending the total corpus isn't possible. Builders layer retrieval pipelines, chunking methods, immediate engineering methods, and multi-agent orchestration programs on high of fashions — all to route across the basic constraint that the mannequin itself can't effectively course of every little thing without delay.
Subquadratic's argument is that these workarounds are costly, brittle, and finally limiting. As CTO Alexander Whedon informed SiliconANGLE in an interview, "I used to manually curate prompts and retrieval programs and evals and conditional logic to chain collectively the workflows. And I believe that that’s sort of a waste of human intelligence and in addition limiting to the product high quality."
Subquadratic's repair is deceptively easy: cease doing the mathematics that doesn't matter
The corporate's method, referred to as Subquadratic Sparse Consideration or SSA, is constructed on an easy premise: a lot of the token-to-token comparisons in commonplace consideration are wasted compute. As a substitute of evaluating each token to each different token, SSA learns to determine which comparisons really matter and computes consideration solely over these positions. Crucially, the choice is content-dependent — the mannequin decides the place to look based mostly on which means, not on fastened positional patterns. This enables it to retrieve particular info from arbitrary positions throughout a really lengthy context with out paying the quadratic tax.
The sensible payoff scales with context size — precisely the inverse of the issue it's making an attempt to resolve. In response to the corporate's technical weblog, SSA achieves a 7.2x prefill speedup over dense consideration at 128,000 tokens, rising to 52.2x at 1 million tokens. As Whedon put it: "For those who double the enter measurement with quadratic scaling legal guidelines, you want 4 instances the compute; with linear scaling legal guidelines, you want simply twice." The corporate says it educated the mannequin in three phases — pretraining, supervised fine-tuning, and a reinforcement studying stage particularly concentrating on long-context retrieval failures — instructing the mannequin to aggressively use distant context quite than defaulting to close by info, a refined failure mode that quietly degrades efficiency in present programs.
Three benchmarks paint a powerful image, however what they pass over could matter extra
On the floor, SubQ's benchmark numbers are aggressive with or superior to fashions constructed by organizations spending billions of {dollars}. On SWE-Bench Verified, it scored 81.8% in comparison with Opus 4.6's 80.8% and DeepSeek 4.0 Professional's 80.0%. On RULER at 128,000 tokens, a regular benchmark for reasoning over prolonged inputs, SubQ scored 95% — edging out Claude Opus 4.6 at 94.8%. On MRCR v2, a demanding check of multi-hop retrieval throughout lengthy contexts, SubQ posted a third-party verified rating of 65.9%, in contrast with Claude Opus 4.7 at 32.2%, GPT-5.5 at 74%, and Gemini 3.1 Professional at 26.3%.
However a number of particulars warrant scrutiny. The benchmark choice is slender — precisely three checks, all emphasizing long-context retrieval and coding, the exact duties SubQ is designed for. Broader evaluations throughout common reasoning, math, multilingual efficiency, and security haven’t been revealed. The corporate says a complete mannequin card is "coming quickly."
In response to The New Stack, every benchmark mannequin was run solely as soon as resulting from excessive inference price, and the SWE-Bench margin is, as the corporate's personal paper acknowledges, "harness as a lot as mannequin." In benchmark methodology, single runs with out confidence intervals depart room for variance. There’s additionally a big hole between SubQ's analysis outcomes and its manufacturing mannequin. On MRCR v2, the corporate reported a analysis rating of 83 — however the third-party verified manufacturing mannequin scored 65.9. That 17-point hole between the lab end result and the delivery product is notable and largely unexplained.
Subquadratic additionally informed SiliconANGLE that on the RULER 128K benchmark, SubQ scored 95% accuracy at a value of $8, in contrast with 94% accuracy and about $2,600 for Claude Opus — a outstanding price declare. However the firm has not publicly disclosed particular API pricing, making it unimaginable to independently confirm the cost-per-task comparisons.
The AI analysis neighborhood's verdict ranges from 'real breakthrough' to 'AI Theranos'
Inside hours of the announcement, the AI analysis neighborhood erupted right into a debate that crystallized round a single query: Is that this actual?
AI commentator Dan McAteer captured the binary temper in a broadly shared publish: "SubQ is both the largest breakthrough for the reason that Transformer… or it's AI Theranos." The comparability to the notorious blood-testing fraud firm could also be unfair, nevertheless it displays the dimensions of the claims being made. Skeptics zeroed in on a number of strain factors. Distinguished AI engineer Will Depue initially famous that SubQ is "nearly certainly a sparse consideration finetune of Kimi or DeepSeek," referring to present open-source fashions.
Whedon confirmed this on X, writing that the corporate is "utilizing weights from open-source fashions as a place to begin, as a operate of our funding and maturity as an organization." Depue later escalated his criticism, writing that the corporate's O(n) scaling claims and the speedup numbers "don't appear to line up" and referred to as the communication "both extremely poorly communicated or simply not actual."
Others raised structural questions. One developer famous that if SubQ actually reduces compute by 1,000x and prices lower than 5% of Opus, the corporate should not have any hassle serving it at scale — so why gate entry via an early-access program? Developer Stepan Goncharov referred to as the benchmarks "very attention-grabbing cherry-picked benchmarks," whereas one other commenter described them as "suspiciously good."
However not everybody was dismissive. AI researcher John Rysana pushed again on the Theranos framing, writing that the work is "simply subquadratic consideration performed properly which could be very significant for lengthy context workloads," and that "odds of it being BS are extraordinarily low." Linus Ekenstam, a tech commentator, mentioned he was "extraordinarily intrigued to see the real-world implications" significantly for complicated AI-powered software program.
Magic.dev made strikingly comparable claims two years in the past — after which went quiet
Maybe probably the most pointed critique of SubQ's launch comes not from its particular claims however from latest historical past. Magic.dev introduced a 100-million-token context-window mannequin in August 2024, with a claimed 1,000x effectivity benefit, and raised roughly $500 million on the energy of these claims. As of early 2026, there isn’t a public proof of LTM-2-mini getting used outdoors Magic.
The parallels are uncomfortable. Each corporations claimed large context home windows. Each touted roughly 1,000x effectivity positive aspects. Each focused software program engineering as their main use case. And each launched with restricted exterior entry.
The broader analysis panorama reinforces the warning. Kimi Linear, DeepSeek Sparse Consideration, Mamba, and RWKV all promised subquadratic scaling, and all confronted the identical downside: architectures that obtain linear complexity in concept usually underperform quadratic consideration on downstream benchmarks at frontier scale, or they find yourself hybrid — mixing subquadratic layers with commonplace consideration and dropping the pure scaling advantages.
A broadly cited LessWrong evaluation argued that these approaches "are all higher considered 'incremental enchancment quantity 93595 to the transformer structure'" as a result of sensible implementations stay quadratic and "solely enhance consideration by a continuing issue."
Subquadratic is straight conscious of this historical past. Its personal technical weblog particularly addresses every prior method — fixed-pattern sparse consideration, state house fashions, hybrid architectures, and DeepSeek Sparse Consideration — and argues that SSA avoids their tradeoffs. Whether or not it really does stays an empirical query that solely unbiased analysis can settle.
A five-time founder, a former Meta engineer, and $29 million to show the doubters improper
The staff behind the claims issues in evaluating them. CEO Justin Dangel is a five-time founder and CEO with a observe document throughout well being tech, insurancetech, and client items, and his corporations have scaled to a whole bunch of staff, attracted institutional backing, and reached liquidity. CTO Alexander Whedon beforehand labored as a software program engineer at Meta and served as Head of Generative AI at TribeAI, the place he led over 40 enterprise AI implementations.
The staff consists of 11 PhD researchers with backgrounds from Meta, Google, Oxford, Cambridge, ByteDance, and Adobe. That may be a credible assortment of expertise for an architecture-level analysis effort. However neither co-founder has revealed foundational AI analysis, and the corporate has not but launched a peer-reviewed paper. The technical report is listed as "coming quickly."
The funding profile is uncommon for an organization making frontier AI claims. Subquadratic raised $29 million at a reported $500 million valuation — a steep value for a seed-stage firm with no publicly accessible mannequin, no peer-reviewed analysis, and no disclosed income. The investor base, led by Tinder co-founder Mateen and former SoftBank accomplice Villamizar, skews towards client tech and progress investing quite than deep technical AI analysis. The corporate isn’t open-sourcing its weights however plans to supply coaching instruments for enterprises to do their very own post-training, and has set a 50-million-token context window goal for This autumn.
The actual check for SubQ isn't benchmarks — it's whether or not the mathematics survives unbiased scrutiny
Strip away the advertising language and the social media drama, and the underlying query Subquadratic is asking is genuinely essential: Can AI programs break freed from quadratic scaling with out sacrificing the standard that makes them helpful?
The stakes are monumental. If consideration might be made actually linear with out degrading retrieval and reasoning, the economics of AI shift essentially. Enterprise functions that at this time require elaborate retrieval pipelines — processing complete codebases, contracts, regulatory filings, medical data — grow to be single-pass operations. The billions of {dollars} presently spent on RAG infrastructure, context administration, and agentic orchestration grow to be partially redundant.
Whedon's willingness to have interaction publicly with technical criticism — posting a technical weblog inside hours of pushback — suggests a staff that understands it wants to indicate its work, not simply describe it. And to its credit score, the corporate acknowledged brazenly that it builds on open-source foundations and that its mannequin is smaller than these on the main labs.
Each frontier mannequin in 2026 advertises a context window of not less than 1,000,000 tokens, however nearly none of them are literally nice at making use of all that info. The hole between a nominal context window and a useful one — between what a mannequin accepts and what it reliably causes over — stays some of the essential unsolved issues in AI. Subquadratic says it has closed that hole. If unbiased analysis confirms that declare, the implications would ripple far past a single startup's valuation. If it doesn't, the corporate joins a rising checklist of long-context guarantees that sounded revolutionary on launch day and unremarkable six months later.
In computing, each basic constraint finally falls. When it does, the breakthrough by no means comes from the route the business anticipated. The query hanging over Subquadratic is whether or not a staff of 11 PhDs and a $29 million seed spherical really discovered the reply that has eluded organizations spending hundreds of instances extra — or whether or not they simply discovered a greater strategy to describe the issue.
