Developer docs
API playgroundTry for free, no card

Search company profiles

Subquadratic

Full company profile

uuid00vqmpw

Namestring
Subquadratic
Legal namestring
Subquadratic Inc.
Websiteurl
subq.ai
Company typeenum
Private
Founded yearint
2026
Descriptiontext

Subquadratic is a Miami-based AI company that builds SubQ, a long-context large language model constructed on a proprietary Subquadratic Sparse Attention (SSA) architecture. SSA replaces the standard O(n²) all-pairs dense attention with content-dependent selection that scales linearly (O(n)) with context length, enabling a native 12-million-token context window. The model is delivered through three product surfaces: the SubQ API (an OpenAI-compatible, full-context API for developers and enterprise teams), SubQ Code (a CLI coding agent that loads entire codebases into a single context and integrates with Claude Code, Codex, and Cursor), and SubQ Search (a long-context deep-research tool). All products are currently in private beta and gated behind a public waitlist.

SubQ was built by replacing dense attention in an existing open-weight frontier model with SSA, then running staged context extension (262K to 512K to 1M to 2M) followed by approximately one trillion tokens of continued pretraining on naturally long artifacts such as books, documents, and repository-scale code. According to third-party validation by Appen, SubQ 1.1 Small achieves 56x faster prefill than FlashAttention-2 at 1M-token context, 64.5x less compute than dense attention, near-perfect needle-in-a-haystack retrieval (100% at 1M to 2M, 98% at 6M to 12M), 99.12% on RULER 128K, 86.2% on MRCR v2, 89.7% pass@4 on LiveCodeBench v6, and 85.4% on GPQA Diamond. Public benchmarks are independently validated through Appen and, going forward, through a Stratix partnership with LayerLens.

The company operates a usage-based API, individual subscription accounts, and enterprise or design-partner contracts under an API Services Agreement, with no public price list disclosed. Founded circa 2025 to early 2026, Subquadratic emerged from stealth in May 2026 alongside a $29 million seed round at a rumored $500 million valuation, backed by angels including Tinder co-founder Justin Mateen (JAM Fund), Javier Villamizar, Grant Gittlin (Lasagna), and Jaclyn Rice Nelson (Coalition Operators), alongside early-stage investors in Anthropic, OpenAI, Stripe, and Brex. The 13-person team includes 11 PhD researchers from Meta, Google, Oxford, Cambridge, BYU, ByteDance, Adobe, and Microsoft, led by CEO Justin Dangel (a five-time founder) and CTO Alex Whedon (formerly Meta and Head of Generative AI at TribeAI). The company is incorporated in Delaware with operational headquarters in Miami and is pre-revenue at the time of analysis.

Short descriptiontext

Subquadratic is a Miami-based AI company building SubQ, a long-context large language model using its proprietary Subquadratic Sparse Attention architecture. It serves developers, enterprise teams, and coding-agent builders via an OpenAI-compatible API and CLI with up to 12M-token context.

Operating statusenum
Operating
Ownership categoryenum
Headcount rangeband
11–50
akta.pro rankint
HeadquartersSan Francisco, United States
HQ citystring
San Francisco
HQ countrystring
United States
HQ regionstring
North America
Markets served

Serves global market

Offices2 records

Each record includes

City, Country, Type, Description, Source

Keyword5 values
long-context LLMs, sparse attention AI, AI inference API, coding agent platform, AI infrastructure platform
Industry1 code
PRIMARYHDAEAAAK
Search / Index Databases
Hardware, Technology & Telecom/Enterprise Platforms & Databases/Database Management Systems (Relational & NoSQL)
NAICS code2 codes
  • Custom Computer Programming Services541511
  • Web Search Portals and All Other Information Services51929
SIC code2 codes
  • Services-Computer Programming, Data Processing, Etc.7370
  • Services-Prepackaged Software7372
Product category
Large Language Model Infrastructure
Social media profiles2 records
GTM motion4 records

Each record includes

Type, Description, Source

Revenue model3 records
1SubQ API (usage-based)
TypeUsage Based
Description

Full-context API sold to developers and enterprise teams for processing full repositories and pipeline states in a single API call at linear cost. Currently in private beta, with usage-based billing implied per Terms of Use.

subq.ai
2Individual subscriptions
TypeSubscription Recurring
Description

Paid subscription accounts for individual users of Subquadratic's foundational model and speech processing APIs (text-to-speech, speech-to-text, speech-to-speech), with automatic periodic renewal billing.

subq.ai
3Enterprise / design-partner contracts
TypeManaged Services
Description

Enterprise sales motion for design partners and broader enterprise rollout through 2026, delivering SubQ via API Services Agreement with usage-based and contracted pricing.

subq.ai
Marketing channels8 records

Each record includes

Title, Type, Stage, Description, Source

Distribution channels4 records

Each record includes

Title, Type, Scope, Target buyer, Description, Source

Cost components4 values
Technology or R&D, Personnel, Infrastructure, Marketing or Sales
Pricing details2 tiers
1Usage-based API (private beta)
ModelUsage-basedBilling cadencePay-as-you-go
Notes

API billed based on usage metrics (tokens processed, API calls, etc.) per Terms of Use; company claims $8 per task vs ~$2,600 on comparable frontier models, and 'less than 5% of Claude Opus cost for 1M tokens.' No public price list — access via waitlist.

subq.ai
2Individual subscription accounts
ModelSubscriptionBilling cadenceMonthly
Notes

Subscription accounts billed on agreed periodic renewal; usage-based charges for tokens, characters, minutes of audio, or API calls; 30-day notice on price increases; non-refundable payments.

subq.ai
GTM typeB2B
B2B
Offering typeSoftware
Software
Brand1 of 3 records shown
1SubQ
Description

Subquadratic's flagship large language model built on the proprietary Subquadratic Sparse Attention (SSA) architecture, supporting up to 12M-token context windows. Available via API and as a foundation for other products.

subq.ai
+2 more records
Core offering1 text field

Subquadratic builds and sells the SubQ family of large language models, a long-context LLM platform built on its proprietary Subquadratic Sparse Attention (SSA) architecture that replaces O(n²) dense attention with content-dependent sparse attention that scales linearly (O(n)) with context length. SubQ is delivered as a full-context, OpenAI-compatible API for developers and enterprise teams, a CLI coding agent (SubQ Code) that loads entire codebases into a single context window, and a long-context Deep Research tool (SubQ Search), all backed by a native 12-million-token context window.

Differentiator
Functional benefit
Problem solved
Quantifiable outcome1 of 8 values shown
  • 56× faster prefill than FlashAttention-2 at 1M-token context (third-party verified by Appen)
+7 more records
Product overview1 text field

Subquadratic offers a unified long-context AI platform anchored on its proprietary SubQ large language model and built around a single core innovation: Subquadratic Sparse Attention (SSA), a linearly scaling attention architecture that replaces standard dense attention. The core SubQ model is delivered through three product surfaces — the SubQ API (a full-context, OpenAI-compatible API for developers and enterprise teams), SubQ Code (a CLI coding agent that loads entire codebases into one context and plugs into Claude Code, Codex, and Cursor), and SubQ Search (a long-context Deep Research tool) — all of which share the same 12M-token context window and linear-cost economics. The current model lineup is led by SubQ 1.1 Small (the latest iteration in private design-partner deployment), which sits alongside the earlier SubQ 1M-Preview preview release, with a broader 2M–12M model family planned for end of 2026.

Product and service3 records
1SubQ API
CategoryAI Inference API
Description

Full-context, OpenAI-compatible API for developers and enterprise teams that processes full repositories and pipeline states in a single API call at linear cost, with a 12M-token context window, streaming, and tool use.

2SubQ Code
CategoryCoding Agent
Description

Coding agent CLI built on SubQ that loads entire codebases into a single context window, enabling plan/execute/review across a full repository without multi-agent orchestration; plugs into Claude Code, Codex, and Cursor.

3SubQ Search
Scale indicator11 records

Each record includes

Type, Value, Description, Source

Partnership5 partners
Strategic tierCoreTypeStrategic or Co-development PartnerAnnounced on2026-05-14
Description

Subquadratic partnered with LayerLens to evaluate SubQ on the Stratix benchmark platform covering 200+ models and ~100 benchmarks. Long-context retrieval, positional consistency, synthesis, reasoning, coding, instruction-following, and tool-use evaluations will be run, with public results published at stratix.layerlens.ai and per-benchmark breakdowns, methodology, strengths, and limitations. Subquadratic is adopting Stratix Enterprise for all future SubQ releases, creating a continuous independent evaluation layer.

Strategic tierCoreTypeStrategic or Co-development Partner
Description

Appen independently verified SubQ's benchmark results (third-party-validated RULER, MRCR v2, LiveCodeBench, and SubQ 1.1 Small efficiency numbers), publishing a public whitepaper at appen.com/whitepapers/subquadratic-preview-model-benchmark-evaluation. The validation is referenced throughout Subquadratic's product materials as evidence for performance claims.

3Claude Code (Anthropic)
Strategic tierMinorTypeOEM/ Whitelabel/ Licensing Partner
Description

SubQ Code is designed to plug into Claude Code, Codex, and Cursor as a long-context layer that auto-redirects expensive model turns. This is a documented integration rather than a formal commercial partnership announcement.

subq.ai
4Codex (OpenAI)
Strategic tierMinorTypeOEM/ Whitelabel/ Licensing Partner
Description

SubQ Code plugs into Codex as part of its supported integration list for long-context coding tasks.

subq.ai
Strategic tierMinorTypeOEM/ Whitelabel/ Licensing Partner
Description

SubQ Code plugs into Cursor as part of its supported integration list for long-context coding tasks.

Recent move5 records

Each record includes

Date, Type, Title, Description, Source

Expansion highlight6 records

Each record includes

Type, Description

Peers10 records
TypeBroad incumbent
Description

Anthropic's Claude family is a frontier LLM with industry-leading long-context support (1M tokens in Claude Sonnet 4) and is one of SubQ Code's explicit integration targets. As a broad incumbent, it offers long-context as one capability within a wider general-purpose model portfolio rather than specializing in the architecture-level cost niche Subquadratic targets.

TypeDirect peer
Description

Mistral builds efficient open-weight foundation models (Mistral Large, Codestral, Pixtral) with strong long-context performance. It is comparable as a foundation model competitor pursuing architectural efficiency, offering enterprise APIs, and serving coding-agent workloads that overlap with SubQ Code.

TypeDirect peer
Description

AI21 Labs builds enterprise foundation models, most notably the Jamba family, which combines Mamba state-space layers with Transformer attention for long-context efficiency. It is directly comparable to Subquadratic because both are pursuing hybrid/efficient attention architectures as alternatives to dense transformers.

TypeDirect peer
Description

Inflection AI builds and fine-tunes large foundation models for enterprise and consumer applications with a focus on differentiated architecture and capabilities. Comparable to Subquadratic as a foundation model lab pursuing a non-commodity model architecture rather than fine-tuning existing open weights.

TypeDirect peer
Description

DeepSeek builds efficient open-weight foundation models (DeepSeek-V2/V3) with innovations like Multi-head Latent Attention (MLA) and DeepSeekMoE targeting long-context efficiency. It is highly comparable to Subquadratic because both pursue architectural innovation to cut long-context inference cost, and both rely on open-weight lineage (Qwen for SubQ; DeepSeek's own models for ecosystem distribution).

TypeBroad incumbent
Description

OpenAI's GPT-5 and o-series models define the frontier LLM benchmark, and Subquadratic's API is explicitly OpenAI-compatible to ride on this developer mindshare. As a broad incumbent, it competes across the full model stack and ecosystem with vastly greater compute, data, and distribution.

TypeEmerging player
Description

Letta (formerly MemGPT) is an emerging player focused on long-memory AI agents and persistent agent state. Comparable to Subquadratic's long-horizon agentic positioning, particularly SubQ Code and SubQ Search, though Letta emphasizes memory-augmentation architectures rather than a new attention mechanism.

TypeDirect peer
Description

Magic is a frontier AI research lab focused on long-context reasoning and autonomous coding agents with substantial venture backing. It is directly comparable to Subquadratic's SubQ Code positioning around full-repository reasoning and long-horizon agentic coding workflows.

TypeBroad incumbent
Description

Google's Gemini family offers multi-million-token context windows and frontier coding/reasoning performance, directly competing with SubQ on long-context workloads. Comparable as a broad incumbent pursuing long-context capabilities with massive compute, native multimodal support, and deep enterprise distribution via Google Cloud.

TypeEmerging player
Description

Contextual AI builds enterprise-focused long-context AI and 'RAG 2.0' systems targeting the same financial, legal, and enterprise-document workflows as Subquadratic. Comparable as an emerging long-context player, though its approach relies on retrieval-augmentation rather than a single multi-million-token model.

Market position
Strengths4 records

Each record includes

Headline, Details, Source

Weaknesses4 records

Each record includes

Headline, Details, Source

Competitive moat5 records

Each record includes

Type, Details

Key risks6 records

Each record includes

Headline, Details, Source

Key highlights7 records

Each record includes

Headline, Details, Source

Customer concentration

Classification, Details

Segment6 records

Each record includes

Title, Type, Primary, Description, Pain point addressed, Use case, Source

Ideal customer profile4 records

Each record includes

Profile, Firmographic size, Sales motion, Sales cycle length, Buying structure, Purchase trigger, Buyer persona, Geography, Industry vertical, Primary use case, Description, Pain points, Evidence proof points, Target buyer

Technology focused
Yes
API detail
Has APIbool
Yes

Docs URL, Description

Integration3 records

Each record includes

Title, Type, Description, Source

AI capability9 records

Each record includes

Type, Description, Source

AI maturity
App detail

Has app

Feature6 records

Each record includes

Title, Differentiator, Description, Source

Core technology
Revenue estimate
Valuation estimate
Number of profiles
Profiles2 records

Each record includes

Name, Designation, Designation category, Overview, Profile commentary, Source

No data
Compliance3 records

Each record includes

Name, Class, Description

Funding overview

Funding stage, Last funding date, Total funding USD

Funding rounds1 record

Each record includes

Round, Amount USD, Date, Pre money valuation, Total investors, Investors, News

Investors1 record

Each record includes

Name, Type, Date of entry, Rounds participated, Website

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

M&A

Each record includes

Name, Acquisition type, Announced date, Completed date, Status, Website, News

Investment

Each record includes

Name, Round, Announced date, Lead investor, Website, News

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Subquadratic

Large Language Model Infrastructuresubq.ai

Subquadratic is a Miami-based AI company building SubQ, a long-context large language model using its proprietary Subquadratic Sparse Attention architecture. It serves developers, enterprise teams, and coding-agent builders via an OpenAI-compatible API and CLI with up to 12M-token context.

What Subquadratic does

Subquadratic is a Miami-based AI company that builds SubQ, a long-context large language model constructed on a proprietary Subquadratic Sparse Attention (SSA) architecture. SSA replaces the standard O(n²) all-pairs dense attention with content-dependent selection that scales linearly (O(n)) with context length, enabling a native 12-million-token context window. The model is delivered through three product surfaces: the SubQ API (an OpenAI-compatible, full-context API for developers and enterprise teams), SubQ Code (a CLI coding agent that loads entire codebases into a single context and integrates with Claude Code, Codex, and Cursor), and SubQ Search (a long-context deep-research tool). All products are currently in private beta and gated behind a public waitlist.

SubQ was built by replacing dense attention in an existing open-weight frontier model with SSA, then running staged context extension (262K to 512K to 1M to 2M) followed by approximately one trillion tokens of continued pretraining on naturally long artifacts such as books, documents, and repository-scale code. According to third-party validation by Appen, SubQ 1.1 Small achieves 56x faster prefill than FlashAttention-2 at 1M-token context, 64.5x less compute than dense attention, near-perfect needle-in-a-haystack retrieval (100% at 1M to 2M, 98% at 6M to 12M), 99.12% on RULER 128K, 86.2% on MRCR v2, 89.7% pass@4 on LiveCodeBench v6, and 85.4% on GPQA Diamond. Public benchmarks are independently validated through Appen and, going forward, through a Stratix partnership with LayerLens.

The company operates a usage-based API, individual subscription accounts, and enterprise or design-partner contracts under an API Services Agreement, with no public price list disclosed. Founded circa 2025 to early 2026, Subquadratic emerged from stealth in May 2026 alongside a $29 million seed round at a rumored $500 million valuation, backed by angels including Tinder co-founder Justin Mateen (JAM Fund), Javier Villamizar, Grant Gittlin (Lasagna), and Jaclyn Rice Nelson (Coalition Operators), alongside early-stage investors in Anthropic, OpenAI, Stripe, and Brex. The 13-person team includes 11 PhD researchers from Meta, Google, Oxford, Cambridge, BYU, ByteDance, Adobe, and Microsoft, led by CEO Justin Dangel (a five-time founder) and CTO Alex Whedon (formerly Meta and Head of Generative AI at TribeAI). The company is incorporated in Delaware with operational headquarters in Miami and is pre-revenue at the time of analysis.

Subquadratic firmographics

Firmographics
Name
Subquadratic
Legal name
Subquadratic Inc.
Website
https://subq.ai
Company type
Private
Founded year
2026
Operating status
Operating
Headcount range
11–50 employees
Short description
Subquadratic is a Miami-based AI company building SubQ, a long-context large language model using its proprietary Subquadratic Sparse Attention architecture. It serves developers, enterprise teams, and coding-agent builders via an OpenAI-compatible API and CLI with up to 12M-token context.
Ownership category
akta.pro rank

Subquadratic industry classification

Industry
Product category
Large Language Model Infrastructure
NAICS
Custom Computer Programming Services (541511), Web Search Portals and All Other Information Services (51929)
SIC
Services-Computer Programming, Data Processing, Etc. (7370), Services-Prepackaged Software (7372)
akta.pro industry
PRIMARYHDAEAAAK
Search / Index Databases
Hardware, Technology & Telecom/Enterprise Platforms & Databases/Database Management Systems (Relational & NoSQL)

Keywords

  • Long-context LLMs
  • Sparse attention AI
  • AI inference API
  • Coding agent platform
  • AI infrastructure platform

Where Subquadratic is headquartered

Location

Headquarters

HQ city
San Francisco
HQ country
United States
HQ region
North America

Offices2 records

Markets served

Subquadratic business model

Business model
GTM type
B2B
Offering type
Software
Cost components
Technology or R&D, Personnel, Infrastructure, Marketing or Sales

Revenue model

  1. SubQ API (usage-based): Full-context API sold to developers and enterprise teams for processing full repositories and pipeline states in a single API call at linear cost. Currently in private beta, with usage-based billing implied per Terms of Use.
  2. Individual subscriptions: Paid subscription accounts for individual users of Subquadratic's foundational model and speech processing APIs (text-to-speech, speech-to-text, speech-to-speech), with automatic periodic renewal billing.
  3. Enterprise / design-partner contracts: Enterprise sales motion for design partners and broader enterprise rollout through 2026, delivering SubQ via API Services Agreement with usage-based and contracted pricing.

Pricing tiers

ModelBillingPrice
Usage-basedPay-as-you-goUsage-based API (private beta)
SubscriptionMonthlyIndividual subscription accounts

Go-to-market motion4 records

Distribution channels4 records

Marketing channels8 records

Subquadratic product offering

Product offering

Core offering

Subquadratic builds and sells the SubQ family of large language models, a long-context LLM platform built on its proprietary Subquadratic Sparse Attention (SSA) architecture that replaces O(n²) dense attention with content-dependent sparse attention that scales linearly (O(n)) with context length. SubQ is delivered as a full-context, OpenAI-compatible API for developers and enterprise teams, a CLI coding agent (SubQ Code) that loads entire codebases into a single context window, and a long-context Deep Research tool (SubQ Search), all backed by a native 12-million-token context window.

Product overview

Subquadratic offers a unified long-context AI platform anchored on its proprietary SubQ large language model and built around a single core innovation: Subquadratic Sparse Attention (SSA), a linearly scaling attention architecture that replaces standard dense attention. The core SubQ model is delivered through three product surfaces — the SubQ API (a full-context, OpenAI-compatible API for developers and enterprise teams), SubQ Code (a CLI coding agent that loads entire codebases into one context and plugs into Claude Code, Codex, and Cursor), and SubQ Search (a long-context Deep Research tool) — all of which share the same 12M-token context window and linear-cost economics. The current model lineup is led by SubQ 1.1 Small (the latest iteration in private design-partner deployment), which sits alongside the earlier SubQ 1M-Preview preview release, with a broader 2M–12M model family planned for end of 2026.

Differentiator

Problem solved

Functional benefit

Brands

  • SubQ: Subquadratic's flagship large language model built on the proprietary Subquadratic Sparse Attention (SSA) architecture, supporting up to 12M-token context windows. Available via API and as a foundation for other products.
  • SubQ Code
  • SubQ Search

Products and services

  • SubQ API Full-context, OpenAI-compatible API for developers and enterprise teams that processes full repositories and pipeline states in a single API call at linear cost, with a 12M-token context window, streaming, and tool use.
  • SubQ Code Coding agent CLI built on SubQ that loads entire codebases into a single context window, enabling plan/execute/review across a full repository without multi-agent orchestration; plugs into Claude Code, Codex, and Cursor.
  • SubQ Search

Quantifiable outcome

  • 56× faster prefill than FlashAttention-2 at 1M-token context (third-party verified by Appen)
  • +7 more outcomes

Companies that use Subquadratic

Customer profile

Segments6 records

Ideal customer profiles4 records

Subquadratic technology and API

Technology

Technology focussed Yes

API detail

Has API
Yes
API docs
API detail

Core technology

AI maturity

App detail

Integration3 records

AI capability9 records

Feature6 records

Subquadratic partnerships and signals

Strategic signal

Partnerships

Five partnerships are on record, tiered core and minor.

  • LayerLenscoreStrategic or Co-development Partner · 14 May 2026Subquadratic partnered with LayerLens to evaluate SubQ on the Stratix benchmark platform covering 200+ models and ~100 benchmarks. Long-context retrieval, positional consistency, synthesis, reasoning, coding, instruction-following, and tool-use evaluations will be run, with public results published at stratix.layerlens.ai and per-benchmark breakdowns, methodology, strengths, and limitations. Subquadratic is adopting Stratix Enterprise for all future SubQ releases, creating a continuous independent evaluation layer.
  • AppencoreStrategic or Co-development PartnerAppen independently verified SubQ's benchmark results (third-party-validated RULER, MRCR v2, LiveCodeBench, and SubQ 1.1 Small efficiency numbers), publishing a public whitepaper at appen.com/whitepapers/subquadratic-preview-model-benchmark-evaluation. The validation is referenced throughout Subquadratic's product materials as evidence for performance claims.
  • Claude Code (Anthropic)minorOEM/ Whitelabel/ Licensing PartnerSubQ Code is designed to plug into Claude Code, Codex, and Cursor as a long-context layer that auto-redirects expensive model turns. This is a documented integration rather than a formal commercial partnership announcement.
  • Codex (OpenAI)minorOEM/ Whitelabel/ Licensing PartnerSubQ Code plugs into Codex as part of its supported integration list for long-context coding tasks.
  • CursorminorOEM/ Whitelabel/ Licensing PartnerSubQ Code plugs into Cursor as part of its supported integration list for long-context coding tasks.

Scale indicators11 records

Recent moves5 records

Expansion highlights6 records

Subquadratic competitors and assessment

Company assessment

Broad incumbents

  • Anthropic: Anthropic's Claude family is a frontier LLM with industry-leading long-context support (1M tokens in Claude Sonnet 4) and is one of SubQ Code's explicit integration targets. As a broad incumbent, it offers long-context as one capability within a wider general-purpose model portfolio rather than specializing in the architecture-level cost niche Subquadratic targets.
  • OpenAI: OpenAI's GPT-5 and o-series models define the frontier LLM benchmark, and Subquadratic's API is explicitly OpenAI-compatible to ride on this developer mindshare. As a broad incumbent, it competes across the full model stack and ecosystem with vastly greater compute, data, and distribution.
  • Google DeepMind: Google's Gemini family offers multi-million-token context windows and frontier coding/reasoning performance, directly competing with SubQ on long-context workloads. Comparable as a broad incumbent pursuing long-context capabilities with massive compute, native multimodal support, and deep enterprise distribution via Google Cloud.

Direct peers

  • Mistral AI: Mistral builds efficient open-weight foundation models (Mistral Large, Codestral, Pixtral) with strong long-context performance. It is comparable as a foundation model competitor pursuing architectural efficiency, offering enterprise APIs, and serving coding-agent workloads that overlap with SubQ Code.
  • AI21 Labs: AI21 Labs builds enterprise foundation models, most notably the Jamba family, which combines Mamba state-space layers with Transformer attention for long-context efficiency. It is directly comparable to Subquadratic because both are pursuing hybrid/efficient attention architectures as alternatives to dense transformers.
  • Inflection AI: Inflection AI builds and fine-tunes large foundation models for enterprise and consumer applications with a focus on differentiated architecture and capabilities. Comparable to Subquadratic as a foundation model lab pursuing a non-commodity model architecture rather than fine-tuning existing open weights.
  • DeepSeek: DeepSeek builds efficient open-weight foundation models (DeepSeek-V2/V3) with innovations like Multi-head Latent Attention (MLA) and DeepSeekMoE targeting long-context efficiency. It is highly comparable to Subquadratic because both pursue architectural innovation to cut long-context inference cost, and both rely on open-weight lineage (Qwen for SubQ; DeepSeek's own models for ecosystem distribution).
  • Magic AI: Magic is a frontier AI research lab focused on long-context reasoning and autonomous coding agents with substantial venture backing. It is directly comparable to Subquadratic's SubQ Code positioning around full-repository reasoning and long-horizon agentic coding workflows.

Emerging players

  • Letta: Letta (formerly MemGPT) is an emerging player focused on long-memory AI agents and persistent agent state. Comparable to Subquadratic's long-horizon agentic positioning, particularly SubQ Code and SubQ Search, though Letta emphasizes memory-augmentation architectures rather than a new attention mechanism.
  • Contextual AI: Contextual AI builds enterprise-focused long-context AI and 'RAG 2.0' systems targeting the same financial, legal, and enterprise-document workflows as Subquadratic. Comparable as an emerging long-context player, though its approach relies on retrieval-augmentation rather than a single multi-million-token model.

Market position

Strengths4 records

Weaknesses4 records

Competitive moat5 records

Key risks6 records

Key highlights7 records

Customer concentration

Subquadratic social profiles

Digital presence

Subquadratic compliance and trust

Trust signal

Compliance3 records

Subquadratic financial estimates

Financial estimate

Revenue estimate

Valuation estimate

Subquadratic leadership team

Management profile

Number of profiles

Profiles2 records

Subquadratic funding detail

Funding detail

Funding overview

Funding rounds1 record

Investors1 record

Funding detail is available on the Subscription and Enterprise plan.Contact sales →

Subquadratic M&A and investment

M&A and investment

M&A

Investments

M&A and investment is available on the Subscription and Enterprise plan.Contact sales →

Frequently asked questions about Subquadratic

What does Subquadratic do?

Subquadratic builds and sells the SubQ family of large language models, a long-context LLM platform built on its proprietary Subquadratic Sparse Attention (SSA) architecture that replaces O(n²) dense attention with content-dependent sparse attention that scales linearly (O(n)) with context length. SubQ is delivered as a full-context, OpenAI-compatible API for developers and enterprise teams, a CLI coding agent (SubQ Code) that loads entire codebases into a single context window, and a long-context Deep Research tool (SubQ Search), all backed by a native 12-million-token context window.

Is Subquadratic a public or private company?

Subquadratic is a private company. It is classified as venture growth investor backed and is currently operating.

When was Subquadratic founded?

Subquadratic was founded in 2026. It employs 11 to 50 people.

Where is Subquadratic based?

Subquadratic is headquartered in San Francisco, United States, in the North America region.

How does Subquadratic make money?

Three revenue lines are on record. SubQ API (usage-based) is the primary driver. The others are individual subscriptions and enterprise / design-partner contracts.

Who are Subquadratic's main competitors?

Broad incumbents on record are Anthropic, OpenAI and Google DeepMind. Direct peers are Mistral AI, AI21 Labs, Inflection AI, DeepSeek and Magic AI. Emerging players are Letta and Contextual AI.

Does Subquadratic have an API?

Yes. The full-context SubQ API for developers and enterprise teams processes full repositories and pipeline states in a single API call at linear cost. Features include a 12M token context window, streaming + tool use, and OpenAI-compatible endpoints. Currently available via private beta (waitlist). Request API access via /request-early-access. Developer documentation is at subq.ai/docs/subq-1-1-small-model-card.pdf.

What industry is Subquadratic in?

Subquadratic's product category is Large Language Model Infrastructure. Its primary akta.pro industry code is HDAEAAAK, Search / Index Databases. Its NAICS code is 541511 and its SIC code is 7370.

Unlock the full company data

50 free credits on sign-up, no credit card required.

Contact sales
Live signals
FinancialContent Business PageSubquadratic Launches SubQ Snap, the First Long-Context Decision Model Purpose-Built for EnterpriseSubquadratic launched SubQ Snap, a multimodal decision model with a 384K token context window, 12x larger than Jev. It scores 96.3 on Decision RULER at 1M tokens and averages 67.6 across 16 Decision Index benchmarks, leading Jev on 13 of 16 tasks. The model is available for limited access.MIT Technology ReviewThese startups are chasing the next big thing in LLMsMultiple AI startups are developing alternative technologies to overcome the computational inefficiencies and architectural limitations of transformer-based large language models, which are becoming a bottleneck as AI demands grow. Four approaches are profiled: sparse attention (Subquadratic's SubQ), power retention (Manifest AI's PowerCoder), liquid neural networks (Liquid AI's hybrid LFMs), and diffusion-based text generation (Inception's Mercury 2), each offering different trade-offs between speed, efficiency, and capability. Industry observers note that transformers, while foundational, represent just one phase in AI development, with significant innovation still needed to achieve more general intelligence at lower computational cost.The New StackWhat comes after attention? This startup says it already knows.Subquadratic published benchmarks for its SubQ 1.1 Small model, which uses sparse attention to handle up to 12 million tokens. It scores near-perfect on needle-in-a-haystack tests and 99.12% on RULER, but lags frontier models on general capability. The company plans a mid-tier SubQ 1.2 Medium model.NateAI startup Subquadratic claims to break through LLM bottlenecks... public evaluation: Nate NewsUS AI startup Subquadratic has unveiled a new model called "SubQ" that claims to solve the core computational bottleneck in Large Language Models by replacing dense attention with sparse attention, reducing computational overhead as text length increases. External evaluation by Appen showed SubQ achieved 56x faster maximum throughput than Flash Attention-based models, 89.7% on the LiveCodeBench coding benchmark, and 98% accuracy on "needle in a haystack" tests at 6M and 12M token contexts—far exceeding current models capped at 1M tokens. However, the company used weights from Chinese open-source model Qwen as a starting point rather than training from scratch, and critics argue this doesn't fully substantiate claims of completely rebuilding the LLM architecture, while unverified cost claims and limited public release leave key questions unresolved.MIT Technology ReviewA startup claims it broke through a bottleneck that’s holding back LLMsMiami-based AI startup Subquadratic emerged from stealth mode claiming to have solved a mathematical bottleneck in large language models by developing SubQ, a new LLM using sparse attention instead of the standard dense attention mechanism. Independent evaluation by third-party firm Appen partially validated the company's claims, showing SubQ was 56 times faster than FlashAttention in benchmarks and scored 89.7% on coding tests, while also demonstrating a 12-million-token context window compared to the typical one-million-token limit in top models. The company says it can run certain tasks for $8 versus the $2,600 cost of comparable models, though experts remain skeptical and note that benchmarks do not fully capture real-world performance capabilities.The SaaS NewsSubquadratic Raises $29M Seed FundingSubquadratic, an AI company based in Miami, raised $29 million in seed funding in May 2026 to develop its SubQ long-context AI model and related tools. The funding will support the expansion of its API and enterprise AI applications utilizing its sparse-attention transformer architecture.SubstackThe Sequence Radar #857: Last Week in AI: Inside the Machine, Outside the Text BoxThis newsletter summarizes key AI developments from the past week, highlighting Anthropic's publication of research on Natural Language Autoencoders, a technique that translates neural network activations into readable text for interpretability, and OpenAI's release of new voice models pushing AI toward native voice interfaces. The week also saw SubQ's controversial announcement of a 12 million-token context window, DeepSeek and Moonshot attracting significant valuation attention, and Sierra raising capital at levels typically associated with much larger AI infrastructure companies.AI CERTsSubquadratic Debate: Machine Learning Faces New Efficiency ClaimsSubquadratic, a machine learning startup that raised a $29 million seed round at a rumored $500 million valuation, released SubQ 1M-Preview claiming its Sparse Attention architecture delivers up to 1,000× compute reduction over existing attention mechanisms at million-token context lengths. The research community has responded with skepticism, noting that the company released no weights or full technical report, making independent verification currently impossible. Academics have cited theoretical hardness proofs suggesting general subquadratic attention may require trading accuracy for speed, while independent labs and journalists are requesting raw logs and reproducible benchmarks before accepting the claims.GIGAZINESubQ, an efficient AI model designed to outperform Claude Opus in processing long contexts, has emerged, capable of handling 12 million tokens and breaking the limits of Transformer.AI development company Subquadratic has announced its SubQ model, which uses a subquadratic architecture instead of the mainstream Transformer approach, enabling it to process up to 12 million tokens with computational complexity that scales linearly rather than quadratically with input size. The SubQ 1M-Preview test model outperforms Claude Opus 4.7 in long-context processing benchmarks while operating 150 times faster per token and reducing costs to one-fifth of existing models. Subquadratic is offering its coding agent SubQ Code and information analysis AI SubQ Search through private beta, with early API access now available.Var IndiaSubquadratic claims its AI model reduces computeMiami-based startup Subquadratic has announced what it claims is the first large language model with fully subquadratic scaling, where compute requirements grow linearly with context length rather than exponentially. The company's SubQ 1M-Preview model reportedly cuts attention compute by nearly 1,000 times at 12 million tokens compared with conventional transformer architectures, potentially significantly reducing the cost of processing large amounts of information with AI. The announcement, which also introduced SubQ Code, SubQ Search, and an API currently in private beta, drew over 12 million views on X and 30,000 waitlist signups within 24 hours, though the claims have sparked intense debate in the AI research community with some viewing it as potentially groundbreaking while others remain skeptical about whether the claims will stand up to scrutiny.