← Back to feed

SubQ has launched the first frontier-scale LLM built entirely on what it calls Sparse-Selective Attention (SSA) — an architecture that replaces standard quadratic attention with selective computation over only the token relationships likely to matter.

The result, the company says, is a 12 million token context window with compute that scales closer to O(n log n) than the O(n²) that makes standard transformers prohibitively expensive at long contexts.

The Architecture

Standard transformer attention computes relationships between every token pair. At 1 million tokens, that’s 1 trillion potential token pairs — most of them irrelevant. A code review does not need every variable declaration to attend to every comment in an unrelated file. A legal contract does not need paragraph-level cross-attention to appendices the model will never reference. SSA learns which relationships are probably important and skips computing the rest.

The performance figures from the launch:

• 52x faster than FlashAttention at 1M token context
• Less than 5% of Claude Opus cost at 1M+ token context
• 12M token context window — roughly 60x the standard 200K context of most frontier models
• Full-repository and persistent-state agent operation without degradation at long contexts

Infrastructure Backing

SubQ has physical compute anchored: Digi Power X signed a GPU infrastructure agreement with SubQ AI for a 40MW purpose-built data center campus in Columbiana, Alabama, targeted for Q1 2027. The 10-year deal is worth $1.1B with extension options up to $2.5B. The infrastructure commitment suggests this is not a demo project.

What Remains Unverified

SubQ has published speed and cost comparisons against standard transformers. What it has not yet published are independent benchmark scores placing the model against GPT-5.5, Claude Opus 4.7, or Gemini 3.1 Pro on standard capability evaluations (SWE-bench, AIME, GPQA, or Arena ELO).

Sub-quadratic and linear attention architectures have a history of trading capability for speed — particularly at shorter contexts where the quadratic cost is manageable and the quality benefits of full attention are highest. Whether SSA closes or widens that quality gap at practical context lengths is the central question for early access users.

Why It Matters

If the benchmarks hold, SSA architecture addresses the single biggest constraint on long-context agentic AI: cost. Running a 1M-token context through Claude Opus at current pricing is approximately $15 per call. At less than 5% of that cost, persistent-state coding agents, full-codebase analysis, and multi-session legal document review become economically viable at production scale.

Early access is at subq.ai. Independent benchmark submissions have not yet appeared on SWE-bench Verified or Terminal-Bench 2.0.