MIT Technology Review · Jun 19, 2026
MIT Technology Review reports Subquadratic’s claims about its SubQ model
Markdown · JSON · Original source
MIT Technology Review reports, on June 19, 2026, that Subquadratic says its SubQ model uses sparse attention to reduce the calculations involved in processing long text. Sparse attention selects some token relationships instead of comparing every token with every other one; Subquadratic says its selection changes with each text. The company argues this could make SubQ faster and cheaper for coding and searching large data sets, while handling longer inputs than most other models. The article reports that Appen’s tests found strong results on coding and long-context retrieval. But a speed test measures theoretical operating speed, not what a model can do, and cover specific conditions rather than a wide range of real tasks. SubQ is not widely available, making its cost claims harder to verify. The article also notes that Subquadratic reused weights from Qwen, which it says cuts across the company’s claim to have reinvented how work. Independent researcher Will Depue says the public evidence does not yet justify the stronger claim that the company solved the quadratic attention bottleneck. The company’s larger ambition is to change how LLMs are built; whether SubQ’s results hold up in wider use remains unknown.
Headline and summary by Escape Meridian, prepared by Luna from the attributed source. A source check is not independent confirmation of the source’s claims.
Original source
MIT Technology Review · Article
Source details
Original source date: Jun 19, 2026 (2026-06-19T10:40:24Z).
Date basis: Date from the publisher’s own records, through its API.
Added to Escape Meridian: 2026-11-19T11:48:10.054Z. First observed: 2026-11-19T08:50:03.689Z. Retrieved: 2026-11-19T08:52:00.726Z.
Referenced by
Recorded citations in public readings. A reference alone does not establish agreement or influence.
No public references recorded here yet.
Limits
Text as retrieved on November 19, 2026; originally published June 19, 2026. The article says SubQ is not yet widely available, limiting outside testing and making its cost claims harder to verify. Appen’s tests covered a handful of standard tests; the article says do not substitute for a wide range of real tasks.
Open question
Will SubQ’s reported performance hold up when more people can test it on a wide range of real tasks?
Source-stated
MIT Technology Review says Subquadratic claims SubQ uses sparse attention, which selects some token relationships rather than comparing every token with every other one.
Instead of multiplying the number assigned to each token by every other number, sparse attention selects just some of the numbers to multiply.
Source-stated
Subquadratic says its method selects which relationships matter on the fly and changes the selection for each piece of text.
the selection is calculated on the fly and differs for each piece of text the model is given.
Source-stated
Appen found SubQ scored 89.7% on LiveCodeBench, which the article describes as a test of performance on competitive coding problems.
On LiveCodeBench, a test that looks at how well models perform on competitive coding problems taken from real contests, SubQ scored 89.7%
Source-stated
The article cautions that benchmarks cover specific conditions and do not replace testing across a wide range of real tasks.
Testing under very specific conditions is not a substitute for running a model on a wide range of real tasks.
Source-stated
The article says SubQ is not yet widely available, making Subquadratic’s cost claims harder to verify.
Subquadratic's claims about cost are harder to verify because SubQ is not yet widely available.
Source-stated
Independent AI researcher Will Depue says the public evidence does not yet justify the stronger claim that Subquadratic solved the quadratic attention bottleneck.
But the public evidence does not yet justify the stronger claim that they have solved the quadratic attention bottleneck.
Source and contributor context
MIT Technology Review · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Editorial provenance
Version 1. Publication ID: 78c509bd-85df-beca-8333-922e505b7455.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-19T11:43:32.759Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 4eea4b177188dbf34539101848cd7433142cf12f7181e16d6208611717f7487f. Source text hash: aef514bd2226e04afd050971bf7f3bcbb9d9bee3df868c690995a194fb0ffb86.
Revision and source-check history
Version 1 · 2026-11-19T11:48:10.054Z
MIT Technology Review reports Subquadratic’s claims about its SubQ model
MIT Technology Review reports, on June 19, 2026, that Subquadratic says its SubQ model uses sparse attention to reduce the calculations involved in processing long text. Sparse attention selects some token relationships instead of comparing every token with every other one; Subquadratic says its selection changes with each text. The company argues this could make SubQ faster and cheaper for coding and searching large data sets, while handling longer inputs than most other models. The article reports that Appen’s tests found strong results on coding and long-context retrieval. But a speed test measures theoretical operating speed, not what a model can do, and cover specific conditions rather than a wide range of real tasks. SubQ is not widely available, making its cost claims harder to verify. The article also notes that Subquadratic reused weights from Qwen, which it says cuts across the company’s claim to have reinvented how work. Independent researcher Will Depue says the public evidence does not yet justify the stronger claim that the company solved the quadratic attention bottleneck. The company’s larger ambition is to change how LLMs are built; whether SubQ’s results hold up in wider use remains unknown.
Original source
MIT Technology Review · Article
Source details
Original source date: Jun 19, 2026 (2026-06-19T10:40:24Z).
Date basis: Date from the publisher’s own records, through its API.
Added to Escape Meridian: 2026-11-19T11:48:10.054Z. First observed: 2026-11-19T08:50:03.689Z. Retrieved: 2026-11-19T08:52:00.726Z.
Limits
Text as retrieved on November 19, 2026; originally published June 19, 2026. The article says SubQ is not yet widely available, limiting outside testing and making its cost claims harder to verify. Appen’s tests covered a handful of standard tests; the article says do not substitute for a wide range of real tasks.
Open question
Will SubQ’s reported performance hold up when more people can test it on a wide range of real tasks?
Source-stated
MIT Technology Review says Subquadratic claims SubQ uses sparse attention, which selects some token relationships rather than comparing every token with every other one.
Instead of multiplying the number assigned to each token by every other number, sparse attention selects just some of the numbers to multiply.
Source-stated
Subquadratic says its method selects which relationships matter on the fly and changes the selection for each piece of text.
the selection is calculated on the fly and differs for each piece of text the model is given.
Source-stated
Appen found SubQ scored 89.7% on LiveCodeBench, which the article describes as a test of performance on competitive coding problems.
On LiveCodeBench, a test that looks at how well models perform on competitive coding problems taken from real contests, SubQ scored 89.7%
Source-stated
The article cautions that benchmarks cover specific conditions and do not replace testing across a wide range of real tasks.
Testing under very specific conditions is not a substitute for running a model on a wide range of real tasks.
Source-stated
The article says SubQ is not yet widely available, making Subquadratic’s cost claims harder to verify.
Subquadratic's claims about cost are harder to verify because SubQ is not yet widely available.
Source-stated
Independent AI researcher Will Depue says the public evidence does not yet justify the stronger claim that Subquadratic solved the quadratic attention bottleneck.
But the public evidence does not yet justify the stronger claim that they have solved the quadratic attention bottleneck.
Source and contributor context
MIT Technology Review · Publisher
The publisher is identified here. No individual author has been inferred.
Source media
Media retrieval status: unchecked.
Terms and references
Reference context selected by Escape Meridian; separate from source evidence. Entity mentions do not establish authorship, involvement or endorsement.
Benchmark
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Large language model
A language model with many learned numerical settings, used to process and generate language.
Editorial provenance
Version 1. Publication ID: 78c509bd-85df-beca-8333-922e505b7455.
Prepared by openai/gpt-6-luna; grounding review by anthropic/claude-opus-5.5.
Source check: uncertain by anthropic/claude-opus-5.5 on 2026-11-19T11:43:32.759Z. This checks fidelity to the source, not whether its claims are independently true.
Editorial selection: accept; reasons: bridge_admitted_by_model, second_opinion_categorical.
Observation: 4eea4b177188dbf34539101848cd7433142cf12f7181e16d6208611717f7487f. Source text hash: aef514bd2226e04afd050971bf7f3bcbb9d9bee3df868c690995a194fb0ffb86.
uncertain · 2026-11-19T11:43:32.759Z · anthropic/claude-opus-5.5
Publication: 78c509bd-85df-beca-8333-922e505b7455
attribution — pass: This is a single MIT Technology Review article. The reading correctly attributes the claims about SubQ to Subquadratic, the test results to Appen, the cautions to the article and the skeptical quote to Will Depue.
But the public evidence does not yet justify the stronger claim that they have solved the quadratic attention bottleneck.
summary — uncertain: The reading is faithful but leaves central terms unexplained for a general reader. The "quadratic attention bottleneck" is the core problem, yet it never says that doubling the text roughly quadruples the computation. Smallest repair: add one plain clause on that quadratic growth. Also add brief glosses for "weights" (values set in training), Qwen (a Chinese open-source model) and long-context retrieval. Token and the wider sparse attention terms are not in the definitions list; suggest term-library additions.
Double the number of words, and you roughly quadruple the number of computations, a rate of increase known as a quadratic expansion.
date — pass: June 19, 2026 matches the UTC day of originalPublished. The page was edited on June 30, 2026, but the reading relies on core article text. It does not use the Deep Dive sidebar of related stories. No hindsight framing appears, and the forecasts remain attributed to the company.
support — pass: Every claim matches an excerpt from the main article body. The caveats about the speed test, benchmarks, availability, cost verification and Qwen weights are supported. The page was edited after publication, but the cited passages are central body text of the piece rather than appended material. No unsupported benefits or costs are implied.
One nagging issue is that Subquadratic reused the weights (values set within a model during training that determine how it will behave) from a version of the Chinese open-source model Qwen to bootstrap SubQ
media — pass: No media was selected, and no verified media candidates were supplied.
Challenge or correct this reading
Send a correction or an opposing source with this reading’s URL. Disagreement and uncertainty can remain unresolved.
A defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition referenceA language model with many learned numerical settings, used to process and generate language.
Definition referenceA defined set of tests used to compare systems. Results depend on the tasks and measures selected.
Definition reference