Skip to content
Stork.AIstork.ai
AI Tool

SubQ Review

SubQ is a Large Language Model (LLM) built on a sub-quadratic sparse attention architecture designed for extreme efficiency and performance on very long context tasks.

shipped Jun 18, 2026freemium
Domain rating39Monthly visits16K/moAI-readablepartial
SubQ — product screenshot

Why it matters

1Features a 12-million-token context window, enabling multi-million token reasoning.
2Utilizes a novel Subquadratic Sparse Attention (SSA) architecture for O(n) attention scaling.
3Claims to be 56 times faster than FlashAttention-2 at a 1M-token context.
4Reported to use 64.5 times less compute than dense attention, achieving a 1,000x reduction in attention compute at 12M tokens.

Stork’s verdict on SubQ

SubQ delivers multi-million token context efficiently, though its published benchmark sets are narrow.

SubQ reviewed by Stork AI · stork.ai/en/subq

Specs

API Docs

View Documentation →

API Available

Yes, public API

overview

What is SubQ?

SubQ is a Large Language Model (technology) tool developed by Subquadratic (technology) that enables Developers, Enterprise teams, Data engineers, Researchers, and Coding agents to process multi-million token contexts efficiently. It utilizes a novel Subquadratic Sparse Attention (SSA) architecture to overcome traditional transformer limitations. This architecture allows SubQ to process significantly longer context windows more efficiently and at a lower cost than conventional transformer models. Its design facilitates reasoning across extensive document collections, entire codebases, and long-horizon agentic tasks without fragmenting information or compromising quality. SubQ emerged from stealth on May 5, 2026, introducing its 12-million-token context window.

features

Key Features of SubQ

SubQ incorporates several technical features designed for efficient and high-performance processing of very long contexts, leveraging its unique architectural foundation.

  • Sub-quadratic sparse attention architecture (O(n) attention) for extreme efficiency.
  • Multi-million token reasoning, supporting context windows up to 12 million tokens.
  • API available with OpenAI-compatible endpoints for broad integration.
  • Auto-redirects expensive model turns, particularly beneficial for coding agents.
  • Achieves 64.5x less compute than dense attention, with a 1,000x reduction in attention compute at 12M tokens.
  • Demonstrates 56x faster processing than FlashAttention-2 at a 1M-token context.
  • Provides near-perfect performance on single-fact retrieval and multi-task retrieval tasks.
  • Capable of processing full repositories and pipeline states within a single API call.
  • Supports streaming and tool use functionalities via its API.

use cases

Who Should Use SubQ?

SubQ is designed for specific professional and technical personas requiring advanced long-context processing capabilities in their AI applications.

  • Developers and Enterprise teams building AI agents that need to work across full repositories, long histories, and persistent state without quality loss.
  • Data engineers and Researchers involved in multi-document analysis, such as ingesting thousands of pages of regulatory filings or medical records to find non-obvious correlations.
  • Coding agents requiring the ability to load entire codebases into a single context window for architecture-level reasoning, cross-file refactoring, and dependency tracing.
  • Professionals in Financial Analysis and Due Diligence who need to reason across entire collections of financial documents without fragmenting information.
  • Legal and Contract Work specialists analyzing lengthy legal documents to understand terms, qualifications, and exceptions distributed across many pages in a single pass.

how to use

How to Use SubQ

SubQ is primarily accessed via its API, which offers OpenAI-compatible endpoints for integration into existing development workflows. Access is currently managed through a waitlist.

  • 1Join the waitlist for early access to the SubQ API, SubQ Code, and SubQ Search.
  • 2Access the SubQ API via OpenAI-compatible endpoints for seamless integration into existing applications.
  • 3Utilize the API to process multi-million token contexts, such as entire code repositories, legal documents, or financial filings.
  • 4Integrate SubQ into AI agent workflows to enable long-horizon tasks requiring persistent state and reasoning over extensive histories.
  • 5Refer to the SubQ 1.1 Small model card, available via the API documentation URL, for detailed technical specifications and usage guidelines.

pricing

SubQ Pricing & Plans

SubQ operates on a freemium model, offering access to its sub-quadratic LLM capabilities. Specific pricing tiers and usage-based costs are not publicly detailed beyond the freemium offering, with access currently limited to a waitlist for early enterprise customers. Subquadratic claims SubQ can run a RULER 128 benchmark for approximately $8, compared to roughly $2,600 for Anthropic's Opus 4.6 at the same context length, indicating a focus on cost efficiency.

  • Freemium access available via waitlist.

Pros

  • +Achieves a 12-million-token context window with high efficiency due to its Subquadratic Sparse Attention (SSA) architecture.
  • +Claims significant cost reductions, with a RULER 128 benchmark run costing approximately $8 compared to $2,600 for Anthropic's Opus 4.6.
  • +Reported to be 56 times faster than FlashAttention-2 at a 1M-token context and uses 64.5 times less compute than dense attention.
  • +Demonstrates high performance on long-context retrieval, scoring over 90% on needle-in-a-haystack tasks at 12M context in research settings.
  • +Achieved strong coding performance, scoring 89.7% on LiveCodeBench and 81.8% on SWE-Bench Verified.
  • +Offers OpenAI-compatible API endpoints, simplifying integration for developers.

Cons

  • −Access is currently limited to early access via a waitlist, restricting broad independent verification of claims.
  • −Published benchmark sets are noted as relatively narrow, potentially focusing on areas where SubQ is expected to excel.
  • −The base model uses weights from existing open-source models (likely DeepSeek V4 family) as a starting point, rather than being trained from scratch.
  • −A notable gap exists between research (83%) and production (65.9%) scores on the MRCR v2 multi-needle retrieval benchmark.
  • −A full peer-reviewed paper or comprehensive model card was not immediately available at launch, leading to calls for more transparency.
  • −Specific pricing details beyond the 'freemium' model are not publicly disclosed.

Similar Tools

SubQ vs Competitors

SubQ differentiates itself in the competitive landscape of Large Language Models by directly addressing the quadratic scaling problem of transformer models through its Subquadratic Sparse Attention (SSA) architecture.

1
Natively Sparse Attention (NSA)↗

NSA enhances long-context modeling in LLMs by seamlessly incorporating sparsity into both training and inference, utilizing hierarchical token modeling and a hardware-aligned design for real-world speedups.

Like SubQ, NSA focuses on sparse attention for long contexts and efficiency. NSA emphasizes hierarchical token modeling and hardware alignment for real-world speedups in both training and inference, aiming to maintain accuracy and contextual relevance.

2
Dynamic Hierarchical Sparse Attention (DHSA)↗

DHSA is a data-driven framework that dynamically predicts attention sparsity online without retraining the base LLM, achieving significant prefill speedups and preserving near-dense accuracy.

DHSA directly competes with SubQ by offering a dynamic approach to sparse attention for long contexts, focusing on adaptability and significant prefill speedups on existing LLMs, even on resource-constrained devices. It aims to overcome limitations of static sparse methods by adapting to task or input variations.

3
SALE (Low-bit Estimation for Efficient Sparse Attention)↗

SALE is a fine-grained sparse attention method that accelerates long-context LLM prefilling using 4-bit quantized query-key products and block-sparse attention, requiring no parameter training.

SALE offers a fine-grained sparse attention method specifically for accelerating the prefilling stage of long-context LLMs, similar to SubQ's efficiency goal, but with a focus on low-bit quantization and seamless integration into existing systems without retraining.

4
SparDA (Sparse Decoupled Attention)↗

SparDA is a decoupled sparse attention architecture that introduces a 'Forecast' projection to predict and prefetch KV blocks, enabling lookahead selection and improving both prefill and decode speed.

SparDA provides a novel decoupled sparse attention architecture that aims for efficiency in both prefill and decode, directly addressing the same performance and long-context challenges as SubQ, but with a unique lookahead selection mechanism and minimal parameter overhead.

See every SubQ alternative, compared→

More on Stork

Related AI Tools

Other tools in this category, matched by shared tags

GPT WatermarkerSurroundersLifeOS AIWulver: Daily Workout StreaksGenifyPlanPulse
Visit SubQ↗
Connect
𝕏
X / Twitterx.com/subquadratic
in
LinkedInwww.linkedin.com/company/subquadratic-ai/
💬
Discorddiscord.gg/Bc9Rrx2x45

Featured in articles