Subquadratic unveils SubQ, first fully subquadratic LLM with 12M-token context
Subquadratic, established by Justin Dangel, has introduced SubQ 1M-Preview, the inaugural large language model featuring a fully subquadratic design that enables linear scaling of compute with context length. This innovative model processes 12 million tokens and reduces attention computation by almost 1,000 times compared to its counterparts. It achieved a score of 95% on the RULER 128K benchmark, equaling Claude Opus 4.6. SubQ's Sparse Attention operates 52 times faster than FlashAttention while using 63% less compute power. It surpassed Claude Opus 4.7, GPT 5.5, and Gemini 3.1 Pro on MRCR v2, earning an 81.8 score on SWE-Bench Verified. Currently, SubQ is in private beta, accessible via API, SubQ Code, and SubQ Search. The company secured $29M in seed funding from investors like Javier Villamizar and Justin Mateen.
Key facts
- SubQ is the first fully subquadratic LLM, with compute scaling linearly with context length.
- SubQ achieves a research result of 12 million tokens, reducing attention compute by nearly 1,000x.
- On RULER 128K, SubQ 1M-Preview scored 95% accuracy, comparable to Claude Opus 4.6 (94.8%).
- SubQ Sparse Attention is 52× faster than FlashAttention with 63% less compute.
- On MRCR v2, SubQ scored 83 (research) and 65.9 (production), outperforming Claude Opus 4.7 (32.2), GPT 5.5 (74), and Gemini 3.1 Pro (26.3).
- On SWE-Bench Verified, SubQ scored 81.8, ahead of Opus 4.6 (80.8) and Deepseek 4.0 Pro (80.0).
- SubQ is available via API, SubQ Code, and SubQ Search in private beta starting today.
- Subquadratic has raised $29M in seed funding from investors including Javier Villamizar, Justin Mateen, Grant Gittlin, and Jaclyn Rice Nelson.
Entities
Artists
- Justin Dangel
- Alex Whedon
- Javier Villamizar
- Justin Mateen
- Grant Gittlin
- Jaclyn Rice Nelson
Institutions
- Subquadratic
- Meta
- Oxford
- Cambridge
- ByteDance
- Adobe
- Microsoft
- Tinder
- JAM Fund
- Lasagna
- Coalition Operators
- Anthropic
- OpenAI
- Stripe
- Brex
- TribeAI