Miami-based Subquadratic raised $29 million in seed funding at a $500 million valuation on May 7, 2026, and launched SubQ 1M-Preview, the first fully subquadratic LLM architecture. SubQ features a 12 million token context window and is 52 times faster than FlashAttention at 1 million tokens while using 63% less compute. The company has not confirmed whether SubQ is built on existing open-source weights or trained from scratch.
Subquadratic selective attention architecture
- ▪At the full 12 million token context window, SubQ's architecture reduces attention compute by nearly 1,000 times compared to standard frontier models
- ▪At 1 million tokens, SubQ's sparse attention mechanism is 52 times faster than FlashAttention while requiring 63% less compute
- ▪SubQ has a research-grade context window of 12 million tokens, which is roughly 9 million words or about 120 books
- ▪SubQ is built on an architecture called Subquadratic Selective Attention (SSA) that processes context linearly rather than quadratically
Benchmark performance results
- ▪Most frontier models top out at around 1 million tokens of context, and even at that length, performance tends to degrade significantly
- ▪On the MRCR v2 benchmark, GPT 5.5 scores 74%, Claude Opus 4.7 scores 32.2%, and Gemini 3.1 Pro scores 26.3%
- ▪SubQ scored 81.8% on SWE-Bench Verified, edging out Opus 4.6 at 80.8% and DeepSeek 4.0 Pro at 80.0%
- ▪SubQ's research model scored 83% on the MRCR v2 benchmark, while its production model scored 65.9%
- ▪SubQ scored 95% accuracy on the RULER 128K long-context benchmark at a cost of roughly $8, compared to Claude Opus's 94% accuracy at approximately $2,600
Community skepticism concerns
- ▪AI commentator Dan McAteer described SubQ as either the biggest breakthrough since the Transformer or AI Theranos
- ▪Subquadratic has neither confirmed nor denied whether SubQ is built on top of existing open-source model weights, and the company has not yet released model weights or a full technical report
Seed funding details
- ▪Subquadratic raised $29 million in seed funding at a reported $500 million valuation
- ▪Subquadratic's seed funding investors include Tinder co-founder Justin Mateen, former SoftBank Vision Fund partner Javier Villamizar, and early backers of Anthropic, OpenAI, Stripe, and Brex
Company background information
- ▪Subquadratic's research team includes 11 PhD researchers with backgrounds from Meta, Google, Oxford, Cambridge, ByteDance, and Adobe
- ▪The article about SubQ was published on May 7, 2026
- ▪Subquadratic was founded in 2024 and was previously called Aldea, initially focusing on voice models before pivoting to attention architecture research
- ▪Subquadratic is led by CEO Justin Dangel and CTO Alexander Whedon
- ▪Subquadratic is a Miami-based AI startup with 13 employees
Enterprise application potential
- ▪Subquadratic has announced three products entering private beta: the SubQ API with the full 12 million token context window, SubQ Code for coding tasks, and SubQ Search for deep research capabilities
- ▪Subquadratic is targeting a 50 million token context window by Q4 2026
Perspective of AI skeptics and technical observers
- ▪AI commentator Dan McAteer described SubQ as either the biggest breakthrough since the Transformer or AI Theranos
Story comments
Loading comments…