This startup develops AI models that process multiple types of data, such as images, text, and audio, specifically for applications in Southeast Asia. Their focus is on creating AI solutions tailored to the unique needs and characteristics of the region.
Funding
Funding not disclosed
Founders
Product
Problem
Many AI solutions lack specific adaptation to the linguistic nuances and cultural contexts of Southeast Asia, leading to suboptimal performance and limited applicability in the region. Existing multilingual models often fail to capture the subtleties of local languages and dialects.
Solution
Mesolitica provides a suite of multimodal AI APIs and open-source tools specifically designed for Southeast Asian languages and contexts. Their offerings include multilingual chat models (MaLLaM), retrieval engines, speech recognition, speech synthesis, and translation services optimized for regional dialects and code-switching. Mesolitica's solutions enable developers to build AI applications that better understand and respond to the diverse linguistic landscape of Southeast Asia. The company also offers enterprise solutions for self-hosting on private networks or clouds.
Target Audience
The primary target audience includes enterprises, developers, and researchers who require AI solutions tailored for Southeast Asian languages, including those working on customer service chatbots, content localization, and speech-based applications.
Features
- MaLLaM: Multilingual large language models pretrained on Malaysian text, available in 1.1B, 3B, and 5B parameter sizes with a 4096 context length.
- Retrieval Engine: End-to-end multilingual retrieval engine optimized for Malaysian languages with low latency and reranking capabilities.
- Speech API: Real-time speech-to-text and speech translation API with speaker diarization, supporting streaming tokens and native Websocket multi-user API.
- Translation API: Translation API supporting 21 target languages, including various Malay dialects, with native code-switching and normalization.
- Open-source toolkits: Malaya (natural language toolkit) and Malaya-Speech (speech toolkit) for Malaysian languages.
- Lightweight continuous batching: OpenAI compatibility using HuggingFace Transformers, supporting Encoder-Decoder, Decoder, and Whisper models.
- Context Parallelism: Support for Blockwise Attention, Ring Attention, and Tree Attention for extended context lengths.