SubQ ships 12M-token context model

This is a serious architecture bet, not just a bigger context-window marketing stunt. If the benchmark claims hold up outside Subquadratic’s own testing, SubQ could make long-context agents less dependent on brittle retrieval pipelines.

• –The 52x FlashAttention speedup claim matters more than the raw 12M-token number, because long context only becomes useful when latency and cost stay sane.
• –SubQ looks strongest for codebases, research corpora, and long-running agent state, where preserving full context is more valuable than narrow chat.
• –The benchmark story is still partly self-reported and internally framed, so independent replication will decide whether this is a real frontier leap or a well-tuned demo.
• –SubQ Code is the product angle to watch: if one model can handle whole repositories in a single pass, it could simplify agent orchestration and reduce retrieval plumbing.
• –The $29M seed gives the company runway, but the hard part is proving that sparse attention can stay reliable as customers push it into production workloads.