AISEAN

Southeast Asian Large Language Models by Alibaba DAMO Academy, released as part of the SEA open-model ecosystem. SeaLLMs v3 (7B) supports a broad suite of SEA languages via continued pre-training and instruction tuning. One of the earliest (Dec 2023) SEA-focused open LLMs alongside SEA-LION, covering Burmese, Khmer, Lao, Malay, Thai, Vietnamese, Indonesian, and more.

First open-source Vietnamese reasoning LLM, released September 2025. Packaged on NVIDIA NIM for single-H100 deployability. Represents Vietnam's next-generation sovereign AI model after PhoGPT's freeze, with built-in chain-of-thought reasoning for Vietnamese.

Southeast Asia's first family of open-source, multilingual, multimodal LLMs. v4 is built on Gemma 3 27B with 128K context, image+text understanding, function calling, and SEA-focused post-training. v4.5 expands agentic capabilities with a speculative decoder for up to 6× efficiency.

Thailand's flagship Thai-optimised LLM series. Typhoon 2.5 focuses on agentic AI with multi-step reasoning, improved function calling, high-throughput inference (3,000+ tokens/sec on H100), and superior Thai fluency. Multimodal variants (audio + vision) also available.

Southeast Asian Languages In One Network — a family of open-source LLMs fine-tuned for SEA languages and cultures. v4 (late 2025) introduced multimodality on Gemma 3 27B; v4.5 (March 2026) adds agent capabilities, a speculative decoder for 6× throughput, SEA-Guard safety models, and SEA-LION-Embedding suite. Supports 11 SEA languages. Part of Singapore's National Multi-Modal LLM Project.

Thailand's leading open-source Thai LLM series. Typhoon 2 features 5 model sizes, multimodal (image + audio) capabilities, and state-of-the-art Thai instruction-following. Typhoon 2.1 Gemma is featured in Google DeepMind's Gemmaverse. Typhoon 2.5 is available via opentyphoon.ai API. Deployed in Thai government services (OPDC partnership).

Indonesia's sovereign open-source LLM, built on SEA-LION with AISG support. Upgraded to 70B parameters in June 2025 with multilingual chat service. Operates across Bahasa Indonesia and 4 local languages. All data and GPU infrastructure stored within Indonesian territory (GPU Merdeka sovereign cloud). Available via sahabat-ai.com and GoPay app.

Vietnam's pioneering open-source generative LLM, pre-trained from scratch on a 102B-token Vietnamese corpus (482 GB cleaned). PhoGPT-4B-Chat is fine-tuned on 70K instruction prompts and 290K conversations. Uses a custom byte-level BPE tokenizer with 20K vocabulary types tailored for Vietnamese morphology.

SEA-LION v4 adds multimodal capability (vision + language) to the family. Reasoning enhanced in v3.5. Part of Singapore's S$70M National Multimodal LLM Programme. Full SEA-LION ecosystem: base LLM, embeddings, safety guard.

SOTA on SEA-HELM multilingual benchmark. Continued pre-training on Llama-3.1-8B and Gemma-2-9B. 200B tokens, 16.8M instruction pairs. NVIDIA-optimized via TensorRT-LLM. MIT license. Widely deployed across ASEAN.

SOTA retrieval embeddings for 10 SEA languages. Tested on SEA-BED (Southeast Asia Embedding Benchmark) with human-curated native data. Sets new records on retrieval, reranking, and semantic textual similarity. Essential for RAG in SEA languages.

Dedicated safety layer for SEA-LION family. Culturally attuned — understands SEA-specific harmful content, cultural sensitivities, and local contexts. Launched Feb 4, 2026. Holistic NLP benchmarks + handcrafted SEA cultural diagnostic tests.

Multimodal Empathetic Reasoning and Learning in One Network. Singapore's first AudioLLM. Trained on 62M multimodal samples, 260K hours of audio. SOTA on Singlish ASR. Fuses MERaLiON-Whisper encoder + SEA-LION text decoder. Free API trial available.

Built on Qwen2.5, continual pre-training on 500B tokens (400B SEA-specific). Sailor2-20B achieves 50-50 win rate vs GPT-4o across SEA languages. 13 SEA languages. Reproducibility cookbook published. Community-driven.

Comprehensive SEA multilingual LLM. SeaLLM-13B outperforms ChatGPT-3.5 on SEA languages. Strong on low-resource languages like Lao and Khmer. SeaLLMs-Audio extension released Mar 2025.

First large audio-language model for SEA. Voice interactions across 5 languages. Built on Qwen2-Audio-7B. SOTA on SeaBench-Audio for Indonesian, Thai, and Vietnamese. Complements SeaLLMs text models.

Thai-English open LLM family from Siam Commercial Bank. Typhoon2-Instruct for chat, Typhoon2-Audio for Thai speech (SOTA). Collaboration with AI Singapore on cross-lingual audio. Most capable open Thai model.

Thai sovereign LLM trained to understand Thai context and culture. Released by NSTDA. NVIDIA subsidiary investment for development. Government-owned, deployed across Thai public sector.

Open-source Thai GPT fine-tuned on Thai instruction datasets. Community-driven. Built on Llama and Mistral. Active development with Thai government and academic support. Most-used open Thai instruction model.

Pre-trained RoBERTa-based Thai language model. Trained on 78.5GB Thai text. Best-performing Thai BERT-class model. Used as backbone for Thai NLP tasks across industry and academia.

Vietnamese-first LLM pre-trained from scratch on 102B Vietnamese tokens. Avoids copyright issues by training on curated corpus. PhoGPT-4B-Chat fine-tuned on 300K+ conversations.

Vietnamese instruction-tuned LLM based on Mistral-7B. Community-built. Strong Vietnamese chat performance. Most downloaded open Vietnamese chat model on HuggingFace.

Compact Vietnamese multimodal LLMs (vision + language). VinAI's efficient models optimized for mobile and edge deployment. Strong performance on Vietnamese visual understanding tasks.

Malaysia's first domestically built sovereign LLM. Reflects Malaysian cultural values and constitutional principles. Launched Aug 2025 by National AI Office. Multilingual for Malaysia's four main language communities.

Family of Malay language models trained on 90B tokens from Malaysian contexts. First significant dedicated Malay LLM. Used in Malaysian government and enterprise. Mesolitica also maintains Malaya NLP library.

Indonesian instruction-tuned LLM. Built on Llama-2-7B fine-tuned with Indonesian instruction datasets. Community-developed to address Indonesian language gap in major LLMs.

Multilingual BERT for Indonesian and regional languages. Covers Javanese and Sundanese — two of the world's most-spoken under-resourced languages. Used in IndoNLU benchmark evaluation.

First fully open-source SEA LLM — all training data and code disclosed. Uses parallel data for continual pre-training of OLMo 2. Addresses LLM data transparency risks. Aimed at reproducibility.