Subquadratic, a Miami-based developer of sub-quadratic large language models, has raised $29M in seed funding from investors including Justin Mateen of Tinder fame and Javier Villamizar. The company launched SubQ, its first model featuring a 12M-token context window powered by Subquadratic Sparse Attention (SSA) architecture that scales linearly with input size. The funding, at a reported $500M post-money valuation, will expand its private beta for API, SubQ Code, and SubQ Search products.
Coding Agents Fuel $242B AI Surge
Subquadratic's raise times with record Q1 2026 AI funding of $242B, 80% of global VC, driven by demand for efficient long-context models. Cognition Labs, maker of Devin coding agent, recently entered funding talks at $25B valuation after a $400M+ round. Magic.dev secured $465M for 100M-token models, but quadratic scaling inflates costs at ultra-long contexts. Subquadratic targets this gap with 1000x attention compute savings, ideal for codebase reasoning.
Transformers Hit Quadratic Wall
Transformer models like those powering ChatGPT and Claude require compute scaling quadratically with context length, quadrupling costs when doubling inputs. Performance degrades around 200K tokens despite 1M+ windows, limiting agents on full repositories or long histories. Current sparse attempts like DeepSeek fall short of frontier accuracy on benchmarks such as SWE-Bench.
SSA Enables Linear Scaling
SubQ uses SSA to focus only on relevant token pairs, achieving sub-quadratic complexity without accuracy loss. It delivers 52x faster inference than FlashAttention at 1M tokens and less than 5% the cost of Claude Opus on RULER benchmark ($8 vs. $2,600 per SiliconANGLE). Benchmarks include SWE-Bench Verified at 81.8%, RULER 95.0% at 128K, and MRCR v2 65.9% via company benchmarks.
As CTO Alexander Whedon explained:
"If you double the input size with quadratic scaling laws, you need four times the compute; with linear scaling laws, you need just twice."
SubQ API offers OpenAI-compatible streaming with tools, while SubQ Code handles full-repo context for agents.
Backers Bet on Architecture Shift
Investors like Mateen (Tinder co-founder) and Villamizar (ex-Stripe, Brex) signal conviction in SubQ's breakthrough beyond transformers. Their profiles suggest growth capital for scaling efficient inference amid rising costs. The round values Subquadratic at $500M post-money despite being out of stealth, reflecting hype around agentic AI.
LLM Market Scales to $150B
The large language model market stands at $10.97B in 2026, projected to reach $149.89B by 2035 at 31.8% CAGR. Mistral AI raised over $1B for efficient open models with smaller contexts. Trends favor sparse architectures as agentic workflows demand million-token reasoning without explosion in inference expenses.
PhD-Heavy Team Drives Innovation
Subquadratic's 35-person team includes 11 PhDs from Meta, Google, Oxford, Cambridge, and BYU. CEO Justin Dangel, a 5x founder, and CTO Whedon (ex-Meta) bring expertise in sparse attention. This research pedigree underpins claims of 1000x efficiency gains over dense attention.
As CEO Dangel noted:
"Transformers defined the last decade of AI, yet one fundamental limitation… compute requirements scale quadratically with context length."
50M Tokens Planned by Q4
Subquadratic eyes a 50M-token window by Q4 2026 per The New Stack. Private beta access grows via subq.ai, targeting developers building persistent agents and multi-modal tools.