技术规格

MaLLaM-3B

MaLLaM-3B (Malaysia Large Language Model) is a foundational 3 billion parameter dense model engineered specifically for the Malaysian linguistic context. Developed from scratch by Malaysia AI and Mesolitica, the model addresses the scarcity of high-quality local language representations by leveraging a curated dataset of 90 billion tokens. This training corpus comprises 349GB of diverse Malaysian digital artifacts, including government documents, local news, literature from the Dewan Bahasa Pustaka, and colloquial social media exchanges. By utilizing a custom-trained Byte Pair Encoding (BPE) tokenizer, the model captures unique Malaysian idioms, slang, and cultural references that are often diluted in English-centric foundational models.

Technically, MaLLaM-3B adopts the Mistral transformer-based decoder-only architecture, which facilitates efficient inference and high performance relative to its parameter count. The model utilizes Grouped-Query Attention (GQA) to optimize the KV cache, thereby reducing memory overhead during sequence generation. It implements the SwiGLU activation function and RMSNorm for stable and accelerated convergence during pre-training. For position encoding, the model employs Rotary Position Embeddings (RoPE), enabling it to maintain precise token relationships within its standard 4096-token context window.

Designed primarily for edge deployment and localized applications, MaLLaM-3B is optimized for environments where low-latency text generation and bilingual proficiency in Bahasa Malaysia and English are required. Its compact architecture makes it suitable for integration into mobile applications, localized chatbots, and on-premise document processing systems. Released under the Apache 2.0 license, the model provides an open-weights foundation for researchers and developers to build downstream tasks such as sentiment analysis, summarization, and instruction-following assistants tailored for the Malaysian demographic.

没有可用的 MaLLaM-3B 评估基准。