Subquadratic is a frontier AI research company launching SubQ, a large language model designed for very long context workloads, alongside SubQ Code, a coding agent built on the same architecture. The product is aimed at developers and AI teams whose agents need to reason across entire codebases, large document corpora, or multi-week task histories without the chunking and retrieval scaffolding that current models require. Where the industry standard is 128,000 tokens for many AI models and up to 1 million tokens for frontier cloud models such as Claude Sonnet 4.7 and Gemini 3.1 Pro, SubQ can manage a context window of up to 12 million tokens, maintain accuracy, increase speed and reduce compute cost.
The technical claim behind the launch is architectural rather than incremental. Co-founders Justin Dangel, who is chief executive, and Chief Technology Officer Alexander Whedon told SiliconANGLE the company settled on a proprietary transformer architecture that implements sparse attention, focused on transitioning from a dense, quadratic-scaling architecture to a sparse, linear one. Instead of comparing every token to every other token, SubQ identifies and computes only the relationships that matter, which the company says cuts attention compute by close to 1,000 times at the 12 million token mark and delivers roughly a 52x speedup over FlashAttention at 1 million tokens. On RULER at 128K, SubQ scores 97.1 against Opus 4.6's 94.8, with reported speedups of 7.2x at 128K and 52.2x at 1M in its benchmarks.
The launch matters now because long-context performance has become the binding constraint for serious agent work, and most current solutions paper over it with RAG and orchestration layers that add latency and bias. The company was previously called Aldea and worked on speech models before pivoting , and has raised $29 million to date at a $500 million valuation from investors including former SoftBank Vision Fund partner Javier Villamizar and Tinder co-founder Justin Mat een. Early access to both SubQ and SubQ Code is open through subq.ai.
Comments (14)
Sign in to join the discussion.
N
Ndidi OkaforMay 5, 2026
12M context is cute but I need to know if I can pipe my entire Confluence into it without legal having a stroke. Where do you sit on SOC2 and data residency in EU?
T
Tomás L.May 5, 2026
52x faster than FlashAttention is the kind of number that either changes everything or quietly disappears from the README in six months. Rooting for the former.
C
Carlos BenavidesMay 5, 2026
Reply guys are arguing about benchmarks while the real question is whether the founder is going to drop a technical blog post or make us reverse engineer it from a podcast.
R
Rajdeep BanerjeeMay 5, 2026
The tweet buried the lede under three bullets and a dash. 'First frontier model with 12M context' should have been line one, not line three.
M
miracleMay 5, 2026
Every long-context demo I've ever seen finds the needle in the haystack and then hallucinates the barn. Show me recall on a 10M token contract review and I'll believe you.
P
Paolo MarchettiMay 5, 2026
What does the API surface look like? Streaming on 12M tokens, rate limits, and is there a self-hosted path or are we all sharing one very tired endpoint?
Y
Yuki TanabeMay 5, 2026
Tagline rewrite, on the house: 'Attention, but it scales.' You can Venmo me later.
F
Fatima Z.May 5, 2026
The launch video pacing is genuinely good. Whoever cut the architecture diagram into the benchmark reveal earned their paycheck this quarter.
S
Sven KarlssonMay 5, 2026
Sub-quadratic sparse attention has been the academic unicorn for years. If this actually ships in production with real users, half of NeurIPS owes you a drink.
K
Kavya IyerMay 5, 2026
First three seconds of the demo video are just a logo fade. You had a 52x number and you opened with vibes.
M
Marc de VriesMay 5, 2026
Retention question nobody asks at launch: how many of those 12M tokens does a user actually fill before churning back to a 200k context model that's good enough?
O
Obi AmadiMay 5, 2026
Counterpoint: most people asking for 12M context have a RAG problem they refuse to solve.
L
Lin ZhaoMay 5, 2026
Waiting for the inevitable $SUBQ token gated inference tier. Don't do it. (Please don't do it.)
H
Hema PrasadMay 5, 2026
A long context window is a longer rope. Whether you climb or hang yourself with it depends on the eval suite.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Subquadratic
The first model built for long‑context tasks
The first model built for long‑context tasks
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Subquadratic is a frontier AI research company launching SubQ, a large language model designed for very long context workloads, alongside SubQ Code, a coding agent built on the same architecture. The product is aimed at developers and AI teams whose agents need to reason across entire codebases, large document corpora, or multi-week task histories without the chunking and retrieval scaffolding that current models require. Where the industry standard is 128,000 tokens for many AI models and up to 1 million tokens for frontier cloud models such as Claude Sonnet 4.7 and Gemini 3.1 Pro, SubQ can manage a context window of up to 12 million tokens, maintain accuracy, increase speed and reduce compute cost. The technical claim behind the launch is architectural rather than incremental. Co-founders Justin Dangel, who is chief executive, and Chief Technology Officer Alexander Whedon told SiliconANGLE the company settled on a proprietary transformer architecture that implements sparse attention, focused on transitioning from a dense, quadratic-scaling architecture to a sparse, linear one. Instead of comparing every token to every other token, SubQ identifies and computes only the relationships that matter, which the company says cuts attention compute by close to 1,000 times at the 12 million token mark and delivers roughly a 52x speedup over FlashAttention at 1 million tokens. On RULER at 128K, SubQ scores 97.1 against Opus 4.6's 94.8, with reported speedups of 7.2x at 128K and 52.2x at 1M in its benchmarks. The launch matters now because long-context performance has become the binding constraint for serious agent work, and most current solutions paper over it with RAG and orchestration layers that add latency and bias. The company was previously called Aldea and worked on speech models before pivoting , and has raised $29 million to date at a $500 million valuation from investors including former SoftBank Vision Fund partner Javier Villamizar and Tinder co-founder Justin Mat een. Early access to both SubQ and SubQ Code is open through subq.ai.
Comments (14)
12M context is cute but I need to know if I can pipe my entire Confluence into it without legal having a stroke. Where do you sit on SOC2 and data residency in EU?
52x faster than FlashAttention is the kind of number that either changes everything or quietly disappears from the README in six months. Rooting for the former.
Reply guys are arguing about benchmarks while the real question is whether the founder is going to drop a technical blog post or make us reverse engineer it from a podcast.
The tweet buried the lede under three bullets and a dash. 'First frontier model with 12M context' should have been line one, not line three.
Every long-context demo I've ever seen finds the needle in the haystack and then hallucinates the barn. Show me recall on a 10M token contract review and I'll believe you.
What does the API surface look like? Streaming on 12M tokens, rate limits, and is there a self-hosted path or are we all sharing one very tired endpoint?
Tagline rewrite, on the house: 'Attention, but it scales.' You can Venmo me later.
The launch video pacing is genuinely good. Whoever cut the architecture diagram into the benchmark reveal earned their paycheck this quarter.
Sub-quadratic sparse attention has been the academic unicorn for years. If this actually ships in production with real users, half of NeurIPS owes you a drink.
First three seconds of the demo video are just a logo fade. You had a 52x number and you opened with vibes.
Retention question nobody asks at launch: how many of those 12M tokens does a user actually fill before churning back to a 200k context model that's good enough?
Counterpoint: most people asking for 12M context have a RAG problem they refuse to solve.
Waiting for the inevitable $SUBQ token gated inference tier. Don't do it. (Please don't do it.)
A long context window is a longer rope. Whether you climb or hang yourself with it depends on the eval suite.