AI Singapore at ATxSingapore 2026: A Full Recap

ByAISG | AI Products

ATxSingapore 2026 (20–21 May) was an important moment for the AI Products team at AI Singapore (AISG). Across two days, we doubled down on our local partnerships to bring SEA-LION to the industry, unveiled the next generation of SEA-LION models, launched Project ATLAS as a global open data platform, and showcased a full suite of open-source tools — from edge deployment and voice pipelines to document intelligence and multilingual evaluation.

Day 1, 20 May: AISG × Temus — Expanded MOU

Group of professionals posing for a photo at a signing ceremony, displaying signed Memorandum of Understanding documents. In the background, a banner with the text 'Unlocking Economic and Societal Value' from Temus.

We opened the week by deepening our partnership with Temus. Building on an existing MOU, the renewed commitment extends into joint prototypes, reusable delivery frameworks, and enterprise deployments that translate nationally developed AI capabilities into real-world operating environments. The partnership will also support multilingual AI use cases and the practical application of Singapore-developed models in production settings — bringing SEA-LION one step closer to the enterprises that need it most.

 “Our partnership with Temus is intended to help bridge that gap — by translating locally anchored model and product capabilities into enterprise use cases that can be governed, evaluated and deployed in practice.” – Dr Leslie Teo, Senior Director of AI Products, AISG

“A capable model is just the start. In regulated settings, you also need sovereignty over your proprietary data, domain-specific context, and the infrastructure to move from prototype to production. That is the gap we want to close together.” – Sutowo Wong, Managing Director, AI and Data, Temus

Day 2, 21 May: Building AI for Our Communities — Useful, Accessible, and Safe

The following day, we hosted our SEA-LION event “Building AI for Our Communities: Useful, Accessible, and Safe” — a half-day showcase of our latest model releases, live demonstrations, a panel discussion, and technical deep-dives across the SEA-LION ecosystem.

Opening: The SEA-LION Vision

Dr Leslie Teo opened by framing the day around SEA-LION’s background and motivation for developing regional AI models. He gave an overview of SEA-LION’s expanding ecosystem — the newly launched v4.5 model family, Project ATLAS, our global open data initiative, and the introduction of new developer tools for Southeast Asia such as Voice pipelines, agentic recipes, document intelligence, RAG embedding models. 

A speaker presents at a conference about SEA-LION, highlighting its models, tools, and platforms, with an audience visible in the background.

SEA-LION v4.5 — Our Latest Developments

Mark Pereira (Head of Partnerships, AI Products) and Dr Ngui Jian Gang (Head of Model Development) gave a lookback at SEA-LION’s progress and showcased our latest model developments and features in the newest model family of SEA-LION v4.5. 

A speaker presenting at a conference on AI, with a slide titled 'SEA-LION: Our progress this year' in the background, showcasing information about embedding suites and use cases in Southeast Asia.

SEA-LION v4.5 was an expansion of the SEA-LION family on the most recent Gemma 4 and Qwen 3.6 state-of-the-art open-source models which enables greater multimodal and  agentic deployment across SEA languages:

The new SEA-LION Speculative Decoder LLM component drew significant interest. Unlike conventional multi-token prediction, it drafts entire blocks of tokens in parallel and verifies them in a single pass — producing multiple tokens in the time ordinarily needed for one, with no impact on output quality. Benchmark results across SEA languages including Burmese confirmed dramatic speedups of up to 6x faster token processing when the Speculative Decoder was added to the base model.

A speaker presenting at a conference titled 'Building AI for Our Communities: Useful, Accessible, and Safe,' with a visual display showing details about 'SEA-LION: The new v4.5' in the background.

The session also covered SEA-LION’s growing ecosystem tools: the SEA-LION Embedding Suite for best-in-class multilingual retrieval and semantic understanding; SEA-Guard, a collection of culturally-aware guardrail models fine-tuned to SEA safety values across text and vision; and SEA-LION powered agentic recipes on OpenClaw and Hermes, positioning SEA-LION as a lightweight, OpenAI-endpoint-compatible drop-in backend for developers building across SEA languages.

Project ATLAS — A Global Open Data Initiative

Harsh Dhand (Director, Research & AI Partnerships APAC, Google), Jessica Tan (ATLAS Platform Lead), Mark Pereira (Head of Partnerships, AI Products) and Sue Tran (Country Director, AI for Vietnam) presented our most ambitious platform announcement to date.

A speaker presenting on stage about Project ATLAS at an event, with a large screen displaying information about Project SEAL-D and Project Aquarion in the background.

Project ATLAS represents a clear progression of our open data initiative for low resource languages: from Project SEAL-D,  to Project Aquarium and with support from Google.org in the form of a USD 1 million funding, to today’s evolution into a globally scaled open data platform. The motivation for an open-data platform for languages is urgent — over 90% of AI training data is in English or high-resource languages, locking communities across the Global South out of AI tools for healthcare, education, economic opportunity and many other benefits from the transformational wave of AI.

Project ATLAS addresses this through three components:

  • ATLAS-Data — hosting, annotating, and licensing multimodal datasets across 20+ languages and 10+ knowledge domains (270+ datasets and counting)
  • ATLAS-Arena — a Language Model Arena for real-user, community-driven model evaluation
  • ATLAS-Evals — an evaluation hub with multilingual, culturally-aware benchmarks
A speaker presenting at a conference about AI for communities, with a screen displaying key points related to language diversity and equity. The audience is seated, attentively listening.

Project ATLAS has a partner network that spans over 20 countries including the Philippines, Indonesia, Vietnam, India, Kenya, Chile, and Rwanda. To support this work, we are partnering with Temasek Trust Foundation Advisors to establish a tailored donor-advised fund for Project ATLAS. This initiative is gaining strong momentum through a growing coalition of funders and ecosystem partners, such as the Gates Foundation and the UNDP Global Centre for Technology, Innovation and Sustainable Development — united by a shared conviction that the Global South deserves representative, high-quality AI built on its own languages, cultures, and data.

Our existing efforts in this initiative across low resource languages was highlighted through our collaboration with AI for Vietnam — covering joint research on language benchmarks and agent productivity, and ecosystem acceleration through hackathons and open data cataloguing.

A female speaker presenting information on AI research for Vietnam, with a slide showing various AI initiatives and tools, at a conference on building accessible and safe AI.

Panel: Local Data, Global Models — Rebalancing the AI Ecosystem

Panel discussion on AI for communities featuring multiple speakers in a conference setting.
Moderator: Darius Liu (Head of Global Strategy, Adaption and Advisor, AI Singapore)
Panellists: Dr William Tjhi (AI Singapore), Dr Shikoh Gitau (Qhala), Dr Sara Hooker (Adaption), Mr Mahmudi Yusbi (ASEAN Foundation)

One of the highlights of the event was the panel discussion which brought together five diverse voices from across the AI ecosystem to challenge the status quo and make the case for community-led, inclusive AI development.

The panel opened with a direct provocation: frontier models are improving fast — so why should anyone still invest in localisation? The answer from across the table was unanimous:  General-purpose models consistently fail at the edges — precisely where the communities that need AI most tend to live. Localised models not only perform better for these edge cases; they unlock impact and opportunities that global models cannot reach.

From there, the conversation moved to the structural realities of SEA’s data landscape: uneven dataset quality, limited local representation in training pipelines, and communities that contribute data without meaningfully benefiting from it. Dr Shikoh drew parallels to the AI landscape in Africa, noting that the challenges of data extraction for low resource languages are a combined struggle across all regions.

The panel closed on a question that cut to the heart of the matter: are enterprise users asking for localisation, or simply better-performing models? The distinction matters — if enterprise demand alone shapes the ecosystem, low-resource communities risk being excluded from AI. The panellists were aligned: regional policy, institutional frameworks, and cross-sector coalitions are the foundations for action to turn good intentions into sustained, equitable outcomes.

Group of five individuals standing together at an event titled 'Building AI for Our Communities.' Features two screens displaying the event details. The atmosphere is professional and focused on AI accessibility.

SEA-LION on the Edge — SEA-LION × Intel

Esther Choa (Head of Product, AI Singapore) and Malcolm Chan (Solution Architect, Intel) presented on deploying SEA-LION locally on Intel AI PCs.

A speaker presenting at a conference, discussing a slide titled 'Closing the last mile to the device' with information about OpenVINO and Intel technologies.

Three constraints make cloud-only AI a poor fit for the region: data sovereignty requirements, unreliable connectivity, and cloud inference costs that scale inefficiently for low resource languages. On-device deployment of AI models addresses all three of these challenges. Using OpenVINO and Intel’s AI Super Builder, SEA-LION models can now run fully on local hardware — enabling AI chat, document RAG embedding, and agentic workflows with zero cloud dependency.

A speaker presenting at an event titled 'Bringing Southeast Asian AI to the Edge' with a slide outlining the significance of edge AI for Southeast Asia. The speaker is holding a microphone and wearing a black polo shirt. In the background, a large banner reads 'Building AI for Our Communities: Useful, Accessible, and Safe'.

We are pleased to share that the Qwen-SEA-LION-v4-32B-IT-OV (4-bit and 8-bit) and SEA-LION-ModernBERT-Embedding-600M-OV are now available on HuggingFace for ease of access to deploying SEA-LION on your PC or laptop. 

Open Source Voice Pipelines

Jonathan Heng (Lead AI Engineer, AISG) presented our team’s work on open-source cascaded voice pipelines for SEA languages to better address our region’s preference for voice communication.

A presenter in a white shirt stands on stage at a conference, discussing a slide titled 'The Open-Source Gap - and How We Intend to Close It'. The slide displays comparisons of various automatic speech recognition models for Malay and Indonesian languages, along with word error rates.

Benchmarking of existing solutions revealed that while Audio-Speech-Recognition (ASR) for regional languages like Malay and Indonesian is reasonably strong across both commercial and open-source models, Text-to-Speech (TTS) quality gaps are significantly larger — with far fewer open-source options available. AI Singapore is actively fine-tuning Qwen3’s TTS 1.7B model for Malay and Indonesian, with early results showing strong performance improvement.

Two voice demos were conducted in Malay and Indonesian, running Qwen3 ASR, SEA-LION v4.5, and a custom fine-tuned TTS model end-to-end to showcase our momentum in developing a solution that supports Southeast Asian voices.

Two releases were also announced:

Stay tuned for our Tamil and broader Southeast Asian language TTS coverage.

Multilingual Translation — LLM-as-a-Judge

Leong Wei Qi (Senior AI Engineer, AISG) addressed the challenge of evaluating machine translation quality for Southeast Asian languages. Despite improving benchmark scores, mistranslated language entities, awkward phrasing, and culturally inappropriate outputs are still a common occurrence.

A presenter discussing 'Measuring the gaps' in AI translation at a conference, with a slide displaying key concepts including Rubric Design, LLM Judge, and Self-improvement.

AI Singapore’s approach pairs the use of LLM-as-a-Judge with linguist-defined rubrics — regional experts encode precise judgment criteria, an LLM applies them at scale, and human-in-the-loop analysis refines the process iteratively. Key gaps identified include poor handling of local language entities, grammatical errors in complex sentences, and unnatural idiomatic translations for our regional languages.

Document Intelligence for SEA

This segment addressed the unique challenges of Optical Character Recognition (OCR) for non-Latin Southeast Asian language scripts — a problem that is exacerbated in frontier models. Wei Qi shared that core difficulties include visually similar character pairs in languages like Thai and Tamil, complex vowel and tone mark positioning in Thai and Vietnamese, and a tendency for open-source models to hallucinate characters in low-resolution or handwritten documents.

AI Singapore’s commitment is to continue building lightweight, open, and accurate OCR models for Southeast Asian languages.

ATxSingapore 2026 was a reminder of what becomes possible when the right partners come together around a shared conviction. Our work in this space remains an open and collaborative effort across all our partners to build AI that correctly represents our diversity and uniqueness.

We are deeply grateful to our collaborators — Adaption, AI for Vietnam, ASEAN Foundation, AWS, Gates Foundation, Google, IMDA, Intel, Qhala, Temus, and Temasek Trust Foundation Advisors, and UNDP — for their continued support, partnership, and belief in what we are building together. Of course, we are also grateful to the National Research Foundation (Singapore) as well as our host institution, the National University of Singapore, and Nanyang Technological University.

And to everyone who joined us on the day – thank you for being part of the community that is building AI for all of us.

Group photo of participants at the 'Building AI for Our Communities' event, featuring people smiling and giving thumbs up, with a presentation screen in the background displaying event details.

Get Involved

Explore our models, tools and platforms, and reach out to us at sealion@aisingapore.org if you would like to collaborate.