MERaLiON / A*STAR IAIC

The research institutions behind MERaLiON

Singapore established the National Science and Technology Board in 1991. It became A*STAR in 2002. This institutional history predates MERaLiON by more than two decades. The programme belongs to a public research organisation rather than a company formed around one model release.

I2R was established in 2002. Its earlier work included communications, signal processing and computing. By 2014, its joint laboratories were developing speech recognition, voice authentication and interpretation systems with industry partners. MERaLiON builds on that research setting, although those earlier systems are not automatically MERaLiON models.

A*STAR announced IAIC in 2026 by combining I2R with the Institute of High Performance Computing. The announcement states that projects from both institutes continue under IAIC. Older model cards and the API website still use I2R. Those credits describe the development history; the current project website names IAIC.

NSCC identifies Su Yi as IAIC's Executive Director. His earlier roles included deputy executive director and strategy leadership at IHPC. This is an institute leadership role. Beh Kian Teik is identified separately as A*STAR's CEO in the government's October 2026 address.

National funding and model development

MERaLiON launched in December 2024 within Singapore's S$70 million national multimodal language-model programme. NRF and IMDA support the research. NSCC supplied ASPIRE 2A+ computing for model training. These roles distinguish the model developer from its funders and computing partners.

The first AudioLLM release credits I2R and collaboration with AI Singapore. Its language component comes from AI Singapore's SEA-LION work, with Google Gemma ancestry. Later releases also adapt upstream models. The speech encoders document training from scratch. The directory names these different development contributions in each model article.

The programme's 2026 news records a MERaLiON 3 showcase and Hokkien speech generation. An August consortium meeting reported 19 members. This is evidence of programme and industry activity, not evidence that each member owns the published model weights.

Partners, rights and service access

The consortium has distinct roles for lead users, technology companies and research organisations. Members can contribute evaluation, data, applications or research resources. Its tuned-model access and commercial licensing arrangements differ by membership role. These arrangements are separate from the terms attached to public weight downloads.

The public model licenses combine a permissive license annex with extra conditions and named upstream terms. Redistribution requires specified notices and license information. The generation 3 license also requires acknowledgement for works developed using a model or its outputs. A bare MIT label would omit these conditions.

Hosted API use has separate terms. They permit processing by A*STAR or authorised service providers and require users to meet the applicable consent and data obligations. Availability and outputs are not guaranteed. The published disclaimer does not establish a fixed data-retention period or a service-level commitment.

Catalog coverage and open questions

This profile covers the 12 distinct model releases in the reviewed MERaLiON weight catalog. Four MLX and GGUF repositories are grouped with their generation 2 parents. It also records the separately trained SpeechEncoder 2 ASR-CTC system, whose current access could not be confirmed. This is a bounded catalog review, not an inventory of all A*STAR research.

The API console exposes transcription, audio chat and speech analysis. Endpoint names and served checkpoint identities are not always the same. The model articles distinguish confirmed downloadable weights from hosted service features. No separate visual model is inferred from broad multimodal descriptions.