Appen Research validates SubQ 1.1 12M-token retrieval

// 50d agoBENCHMARK RESULT

Appen Research validates SubQ 1.1 12M-token retrieval

Appen Research has conducted an independent evaluation of Subquadratic's SubQ 1.1 Small Preview model on Needle In A Haystack (NIAH) benchmarks spanning 1 million to 12 million tokens, as well as on LiveCodeBench. SubQ 1.1 Small Preview is built on a subquadratic sparse attention architecture, which is designed to process massive context windows with significantly lower compute requirements and higher inference speeds compared to standard transformer models.

// ANALYSIS

Third-party validation from Appen is a significant milestone for Subquadratic AI, helping to dispel skepticism around the real-world viability of their non-transformer architecture.

–Independent benchmarks are crucial for validating claims of linear scaling and near-perfect retrieval across multi-million token contexts.
–Demonstrating high-fidelity retrieval up to 12 million tokens positions SubQ as a highly competitive alternative to frontier transformers for large-scale codebase analysis.
–The model's success on LiveCodeBench indicates that efficiency gains do not immediately come at the cost of coding capabilities.

// TAGS

subquadraticsubqappen-researchbenchmarkinglong-contextniahllms

DISCOVERED

50d ago

2026-06-16

PUBLISHED

50d ago

2026-06-16

RELEVANCE

7/ 10

AUTHOR

subquadratic

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

MODEL51m ago

Opus 5 Relaxes Cyber-Classifier Restrictions by 85%

Anthropic's Opus 5 release includes updated cyber-classifiers that are 85% less restrictive than Claude 3.5 Sonnet. This update significantly reduces refusals on offensive-security tasks, signaling a key shift in how AI safety guardrails are balanced against specialized utility for security professionals.

TUTORIAL1h ago

Codex Security details AI code generation risks

As rapid development via AI code generation ("vibe coding") grows increasingly popular, products created without thorough security oversight risk shipping critical vulnerabilities to production. This post presents a hands-on examination of OpenAI Codex and AI-assisted workflows, exploring common security pitfalls in machine-generated code and outlining strategies for auditing applications before release.

UPDATE2h ago

Grok Build 0.2.121 adds turn summaries, extension grouping

xAI has launched version 0.2.121 of Grok Build, bringing key developer experience updates to its workspace interface. The release introduces concise agent turn summaries directly on dashboard rows for better context tracking, alongside alphabetical grouping within the extensions modal to streamline navigation.