SubQ posts Appen-verified 12x context at parity with GPT-4 class models via subquadratic attention

Subquadratic exited stealth claiming replacement of transformer dense attention with a subquadratic operator. Initial self-reported scores drew immediate comparison to Theranos. One month later the firm released third-party Appen results covering latency, energy, and 12x context expansion against reference models from OpenAI, Google DeepMind, and Anthropic.

Appen measured SubQ at equivalent pass@1 on coding suites while processing sequences up to 12 times longer than the 128 k token regime standard in 2025 frontier systems. Compute per token fell by roughly 8x at 512 k length. The numbers track theoretical removal of the quadratic term identified in the 2017 "Attention Is All You Need" paper and subsequent linear-attention variants such as Performer (arXiv:2009.14794).

Standard scaling curves hit memory and energy walls once context exceeds ~200 k tokens; SubQ shifts that wall outward. Production inference clusters sized for 128 k context can now ingest full codebases or regulatory filings without proportional GPU-hour growth. Training runs remain untested at this scale.

Independent replication on open checkpoints and measurement of training FLOPs are required before architecture displacement claims can be accepted. Absent those data, SubQ remains an inference-time optimization rather than a training paradigm shift.