Subquadratic launches SubQ, a large language model with a 12 million token context window using a new sparse-attention architecture, on May 5, 2026.

Subquadratic came out of stealth on May 5, 2026 with $29 million in seed funding. The company, founded by CEO Justin Dangel and CTO Alex Whedon, built SubQ using Subquadratic Sparse Attention (SSA), an architecture that scales linearly with context length instead of quadratically.

SubQ's SSA mechanism makes the model choose which positions in a sequence to attend to rather than comparing every token with every other token. The research model ships with a 12 million token context window; the production API offers 1 million tokens. At 1 million tokens, SubQ runs 52x faster than FlashAttention-2 on Nvidia B200s.

On benchmarks: SubQ scored 95.0% on RULER @ 128K (long-context retrieval), 65.9% on MRCR v2 @ 1M (multi-round retrieval), and 81.8% on SWE-Bench Verified (code tasks). On RULER 128K, Subquadratic reports SubQ achieves 95.0% accuracy at $8 of compute, compared with roughly $2,600 for Claude Opus to reach 94.8%.

Three products launched in private beta as of May 5: SubQ API (OpenAI-compatible endpoints with tool use support), SubQ Code (a CLI coding agent), and SubQ Search (a long-context research tool).