What comes after attention? This startup says it already knows.

Subquadratic, a startup that launched earlier this year with a sparse-attention model supporting 12-million token context windows, has published its first model card and benchmarks for SubQ 1.1 Small. CTO Alex Whedon clarifies the company is not purely a sparse-attention company — it is already developing 'zero attention' architectures that drop the attention mechanism entirely, drawing inspiration from Yann LeCun's world model research. The SubQ 1.1 Small model achieves near-perfect scores on long-context retrieval benchmarks (99.12% on RULER at 128K tokens), claims 64.5x less compute than dense attention at 1 million tokens, and runs 56x faster than FlashAttention-2. General capability lands near mid-tier frontier models. The model was built by replacing dense attention in an existing open-weight model with SSA and running ~1 trillion tokens of long-context pretraining. Current access is limited to enterprise design partners with large AI spend, with a limited individual release planned before general availability. The next model is expected to be mid-tier sized rather than frontier-class.