New Zealand teams publishing AI research.

Most public AI research in New Zealand comes from universities, public science organisations and Māori-led language technologists, with a few companies publishing code. For a business, research is useful as code you can reuse, methods you can trust or benchmarks you can rerun on your own work.

New Zealand AI research teams compared

Levy Street first, then alphabetical. “Releases” means code, data or benchmarks the team itself publishes.

Team profiles

Levy Street

Levy Street is a Wellington AI engineering company whose research feeds the systems it builds. Its open-source work includes World of ClaudeCraft, an MIT-licensed browser MMO with a Gymnasium environment for training reinforcement-learning agents (2,303 GitHub stars and 60 contributors on 19 October 2026), and SPAWN D, a pre-alpha control plane for coding agents under MIT or Apache-2.0. Tiny Shakespeare CLM, a language model written as Python and improved by a Claude agent, has no licence and reports training-set scores only. Three team members co-authored GTM-Bench, a Blackpearl Group benchmark, and LucentBench is described in the whitepaper of Lucent, a Levy Street venture, with no public dataset. None of these papers or benchmarks has been peer reviewed.

AI Researchers Association (AIRA)

AIRA is a not-for-profit incorporated society, set up in 2021 to support AI research in New Zealand. It published the third edition of its strategy whitepaper in October 2026 and keeps a directory of research groups, which helped us find several entries here.

Earth Sciences New Zealand

Earth Sciences New Zealand, formed when GNS Science and NIWA merged on 1 July 2025, announced on 20 October 2026 a Climate Dynamics study that used an in-house generative AI model to produce more than 15,000 simulated years of daily rainfall and temperature at 12 km resolution. The article mentions no code or data release.

Te Hiku Media

Te Hiku Media, a Kaitāia charitable media organisation belonging to five Far North iwi, leads Papa Reo, its language technology programme, with Dragonfly Data Science. Its team published a te reo Māori benchmark paper in the NeurIPS 2024 Datasets and Benchmarks Track; the paper does not release new assets. Its code is under the Kaitiakitanga License, which requires permission for any use and bars commercial use without an explicit grant.

Te Pūnaha Matatini

Te Pūnaha Matatini is a Centre of Research Excellence for complex systems, hosted by the University of Auckland. Its AI project studied how AI interacts with the systems it is used to study, with the aim of more interpretable algorithms; the research listing marks it completed.

University of Auckland

Auckland’s Strong AI Lab works on neuro-symbolic AI and publishes code under MIT and Apache-2.0, such as an emotion recognition toolkit. Its Industrial AI Group released the HA-ViD video dataset under a non-commercial Creative Commons licence. School of Computer Science researchers, with a Massey University co-author, published LIBRA, over 360,000 test cases for measuring language-model bias in a New Zealand context. The university’s proposed Aotearoa Agentic AI Platform received $250,000 of seed funding to develop a full proposal.

University of Canterbury

Canterbury’s Spatial and Image Learning group researches image processing and machine learning, develops prototypes for deployment and invites organisations with an image learning problem to get in touch. Its AI research group page describes the group as newly formed and refers to a 2022 workshop.

University of Otago (CAIPP)

Otago’s Centre for Artificial Intelligence and Public Policy researches AI policy, regulation, ethics and governance, with attention to policy options for New Zealand. It says it advises government ministries. This is policy research, not model building.

University of Waikato

Waikato’s AI Institute, Te Ipu o te Mahara, and its machine learning group maintain Weka (GPL), MOA for data streams (GPL-3.0) and CapyMOA (BSD-3-Clause). Waikato says its tools have more than 10 million downloads, and the institute offers research, advisory and consultancy services.

Victoria University of Wellington

Te Herenga Waka’s Centre for Data Science and Artificial Intelligence runs research projects and commercial work for clients, plus postgraduate supervision. Its Evolutionary Computation and Machine Learning group publishes peer-reviewed papers on genetic programming, feature selection and text mining.

We also checked AI New Zealand’s “NZ AI Benchmark”, which is a self-assessment of an organisation’s AI adoption, not a test of AI systems, so it is not listed.

How to judge AI research before you rely on it

Treat a result as evidence only when you could check it yourself.

• The licence is stated, and it allows your use (commercial, modified, redistributed).
• Scores come from held-out data, not the data the system was tuned on.
• The test set and scoring code are public, or you can rerun the test on your own tasks.
• The paper says who ran the benchmark and whether they also built a competing system.
• It has been peer reviewed, or you know it has not.
• Any Māori data was used with consent and under the terms its kaitiaki set.

Our LLM evaluation service builds these tests on a client’s own documents rather than relying on public scores.

Where AI research claims go wrong

• “Open” without a licence. Code on GitHub with no licence is source-available: you can read it, not reuse it. Our Tiny Shakespeare CLM is in that category.
• Training-set scores. A score measured on the data a system was tuned on overstates what it will do on new work.
• Benchmarks only the operator can run. The GTM-Bench paper, which Levy Street team members co-authored, states that its production data environment “limits immediate reproducibility outside the benchmark operator”, and Blackpearl’s own purpose-built system led its results.
• Non-commercial data in a commercial product. HA-ViD’s CC BY-NC licence and Te Hiku’s Kaitiakitanga License both rule out unapproved commercial use.
• Preprints read as settled. The 2020 temporal probability calibration preprint co-written by Levy Street’s Max Polaczuk has been cited by Google Brain researchers, but it is still not peer reviewed.

Licences, Māori data and funding in New Zealand

Te Mana Raraunga’s Māori Data Sovereignty principles say Māori data should be stored in Aotearoa whenever possible and that free, prior and informed consent should underpin its use. Te Hiku’s licence applies that view to code and models.

For funding, Inland Revenue lists the R&D Tax Incentive as a 15% tax credit on eligible R&D, with a minimum of $50,000 of eligible spending a year. The Government committed up to $70 million over seven years to an AI research platform through the New Zealand Institute for Advanced Technology; as at 20 October 2026 we could not find an announcement naming the selected platform. If you want research turned into a working system, start with a workflow assessment.

Written by the Levy Street team in Wellington. Last updated 19 October 2026.