We build frontier modelsthat scale efficiently
Subquadratic is a frontier AI research company building the most compute-, memory- and sample-efficient algorithms and models.
Introducing SubQ
The best model for data-intensive workloads.
SubQ is the first model built for multi-million token reasoning, allowing enterprises to work across full repositories, financial filings, and contract archives for a fraction of the cost.
Enterprises are spending billions to work around the limitations of today's models.
- Token Cost
- Context Rot
- Latency
- Chunking
- Embedding
- Vector DB
- Context Compression
- Hallucinations
- Eval Patches
- Context Drift
- Alignment Gaps
- Token Cost
- Context Rot
- Latency
- Chunking
- Embedding
- Vector DB
- Context Compression
- Hallucinations
- Eval Patches
- Context Drift
- Alignment Gaps
- Token Cost
- Context Rot
- Latency
- Chunking
- Embedding
- Vector DB
- Context Compression
- Hallucinations
- Eval Patches
- Context Drift
- Alignment Gaps
The data-heavy workloads that become possible.
Whole Artifact Reasoning
Analyze entire datasets at once: an entire GitHub repo, all legal documentation, every vendor contract, or years of SEC filings, without chunking, compression, or context loss.
Long Horizon Agents
Agents that hold the full task in context and remember long sequences of events, collapsing brittle multi-agent orchestration into one coherent run.
Search & Retrieval
Surface material insights that you can trust from across an entire company's IP, code, or documents without risk of hallucination.
All your context.
Always available.
Reason across millions of tokens in one prompt: entire repos, whole artifacts, and long-running agent state, with context to spare.
~ Approximate token counts.
Today's transformer-based LLMs waste compute by processing every possible relationship between tokens, but only a small fraction of these relationships matter.
SubQ is built differently. Using a proprietary algorithm called Subquadratic Sparse Attention (SSA), it isolates the tokens and relationships that matter, ensuring compute is used efficiently.
At 2M tokens, SubQ uses 128x less compute than frontier models.
A leader in long-context retrieval and reasoning tasks.
Multi-Fact Retrieval
99.12%Single Fact Retrieval: 1–2M
100%Single Fact Retrieval: 6–12M
98%| Benchmark | SubQ 1.1 Small |  GPT-5.5 |  Opus 4.8 | Sonnet 4.6 | GPT-5.4-mini | GPT-5.4-nano | Haiku 4.5 |
|---|---|---|---|---|---|---|---|
| Graduate-level scienceGPQA Diamond · pass@1 | 85.4 | 93.2 | 92 | 87.5 | 87.5 | 81.7 | 67.2 |
| Agentic financeAutomationBench | 13% | 18% | 16% | 8% | 0% | n/r | 3% |
| Competitive programmingLiveCodeBench v6 · pass@4 | 89.7 | 92 | 92.2 | 88.9 | 78.6 | 78.2 | 69.7 |
n/r = result not reported by the model provider
Technical reportRequest API Access.
Building a new class of LLMs from the architecture up.
Subquadratic is a frontier AI research company that believes the architecture layer is the highest-leverage surface in AI. While other major labs focus on incremental improvements to Transformer models, we're pushing foundational change at the architecture level to build models and products that scale efficiently.