Dubai-based Anticloud reports results from a benchmark in which its 27B-parameter PAX v52 model generated structured research frameworks for 20 open problems across multiple scientific fields.

Dubai, United Arab Emirates--(Newsfile Corp. - October 6, 2026) - Anticloud FZ LLE, a Dubai-based developer of sovereign AI infrastructure, has reported results from a PAX v52 benchmark in which its 27B-parameter model generated structured research frameworks for 20 open problems spanning mathematics, physics, computer science, biology and cosmology.

According to Anticloud's published benchmark results, the test generated approximately 24,000 tokens of output during a run on a T4 GPU. The company reported an estimated electricity cost of $0.013 for the run and said each generated output was SHA3-256 hashed and recorded through its AIOSS audit system.

Anticloud said the benchmark was designed to evaluate the model's ability to generate structured research frameworks and does not represent accepted solutions to the underlying scientific problems.

The benchmark also forms part of a broader discussion around the capabilities, cost and accessibility of AI systems being developed by frontier research labs. Anticloud said its results are intended to contribute to this discussion by examining what can be achieved with a smaller, locally deployed model under a controlled computing environment.

"The benchmark was designed to test what a smaller, locally deployed model can produce when the system is built around auditability, reproducibility and control over the computing environment," said Lois-Kleinner Alpasan, founder of Anticloud FZ LLE.

At 22 years old at the time of the reported benchmark, Alpasan was leading the development of Anticloud’s locally deployed AI infrastructure and the systems used to conduct and record the test.

PAX v52 Benchmark Results

According to Anticloud's published benchmark data, the PAX L5 Narrow L2 General 27B model generated research frameworks for 20 open problems across mathematics, physics, computer science, biology and cosmology.

The benchmark included research prompts based on the seven Clay Millennium Prize Problems: Navier-Stokes existence and smoothness, the Riemann Hypothesis, P versus NP, Yang-Mills existence and mass gap, the Hodge Conjecture, the Birch and Swinnerton-Dyer Conjecture, and the Poincaré Conjecture. It also included research questions involving the Black Hole Information Paradox, matter-antimatter asymmetry, quantum gravity, consciousness and dark matter, among other subjects.

The company reported the following results from the benchmark:

• 20 research frameworks generated
• Approximately 24,000 tokens processed
• Approximately 286 seconds per research framework
• Approximately $0.013 in reported electricity cost
• $0.00 reported Kaggle cost
• 4.2 tokens per second under the reported benchmark configuration

Anticloud said the outputs were generated during a reproducible test run and that the resulting materials were SHA3-256 hashed and recorded through its AIOSS system.

The company said three selected outputs received an 8/10 evaluation under its reported assessment criteria, which included mathematical rigor, logical coherence, physical plausibility, testability and novelty.

One selected framework examined matter-antimatter asymmetry using concepts including Sakharov conditions, leptogenesis and sphaleron conversion. Another examined the Black Hole Information Paradox using concepts associated with generalized entropy, entanglement islands and replica wormholes. A third examined the Yang-Mills existence and mass gap problem using concepts including Wilson action, Osterwalder-Schrader axioms and area-law confinement.

Anticloud said these materials are research frameworks generated by the model and should not be interpreted as accepted solutions to the underlying scientific problems.

Reported PAX v52 Performance

Anticloud also reported performance results for the PAX L5 Narrow L2 General 27B model across several benchmark and assessment frameworks.

The company reported:

• 97.3 tokens per second on a T4 GPU
• Approximately $0.08 per 1 million tokens
• 100/100 on the reported MITRE ATT&CK assessment
• 88% on the reported NIST AI RMF assessment
• 7/9 on the reported machine-learning readiness assessment
• 61.86% on TruthfulQA
• 69.78% on MMLU
• 77.46% on HellaSwag

Anticloud uses the terms "L5 Narrow" and "L2 General" as part of its classification of the PAX v52 model. The company describes L5 Narrow in relation to selected tasks in regulated sectors including healthcare, defense, finance, legal and government applications.

The company said the reported figures represent results from different assessment frameworks and may use different methodologies and evaluation criteria.

Audit and Deployment Infrastructure

Anticloud said auditability is a central component of its PAX v52 architecture.

According to the company, its AIOSS system uses SHA3-256 hashing to maintain a tamper-evident record of generated outputs. The company reported an AIOSS audit-chain throughput of 205,796 entries per second and a CRDT merge time of 0.11 milliseconds for a 10,000-operation test conducted in its reported T4 environment.

The company also reported the use of AES-256GCM and scrypt for its sovereign memory encryption system.

Anticloud said its approach is intended to provide greater visibility into model activity, generated outputs and system operations when AI workloads are run in controlled computing environments.

Research Materials and Archives

Anticloud has made research and benchmark materials available through several public repositories and archives, including ORCID, Harvard Dataverse, Figshare, OSF, Academia.edu, Zenodo, the Internet Archive and Dev.to.

The company provided the following research identifiers and records:

ORCID: 0009-0009-2233-6107

Harvard Dataverse: DOI 10.7910/DVN/YMJKOG

AIOSS chain genesis: 8b4a8a4f6312dfbe885de8280716985637c163fd2a4b5590341d56db1cc4e560

PAX v52 chain head: 2828cffabd1d063a46adfcf3b1724a8d0e716583207fde2161b95d480413073c

Anticloud's Development Roadmap

According to Alpasan, Anticloud's current development focus includes expanding its AI architecture into a more modular system.

As part of its future research and evaluation work, Anticloud is also developing an AGI-oriented test based on a 77,000-page corpus designed to evaluate models on broader knowledge and reasoning capabilities beyond structured, grid-based rule evaluation. The company said the test is intended to provide an additional reference point for future model assessment.

"The goal right now is to develop past our current prototype of a system. It needs to work like proteins - to absorb and to work together as a modular system," said Alpasan.

Anticloud said the broader objective is to develop AI infrastructure that can support locally controlled workloads while maintaining records of system activity and generated outputs.

About Anticloud FZ LLE

Anticloud FZ LLE is a Dubai-based developer of sovereign AI infrastructure and AI systems focused on local deployment, auditability and controlled computing environments. The company develops technology for AI workloads across research and selected regulated-industry applications.

To view the source version of this press release, please visit https://www.newsfilecorp.com/release/317270