Agnes-AI/Agnes-SeaLLM-8b

Agnes-SeaLLM-8B: Compact, Culturally Aware, and High-Performing

Agnes-SeaLLM-8B is an 8 billion parameter Large Language Model (LLM) from Agnes-AI, designed for efficient deployment and superior performance in Southeast Asian languages. It boasts a 32768 token context length and is meticulously tuned to reduce hallucinations and ensure culturally sensitive responses, while maintaining strong performance in English and Chinese.

Key Capabilities & Differentiators

• Compact Efficiency: Enables high-speed inference and low-resource deployment, making it ideal for edge devices.
• Top-Tier Performance: Outperforms comparable open-source models across academic examinations, complex instruction following, mathematics, and high-precision translation.
• Superior Instruction Following: Excels in multi-turn dialogues and executing nuanced tasks with high fidelity.
• Culturally Aware & Reliable: Engineered for reduced hallucinations and increased sensitivity to Southeast Asian cultural nuances.
• Balanced Multilingual Mastery: Achieves consistent, high-quality output across a broad linguistic spectrum, avoiding the "seesaw effect" common in regional models.

Performance Highlights

Agnes-SeaLLM-8B achieves an average score of 75.32 on SeaExam and 74.13 on MMLU, surpassing 8B peers and even outperforming larger models like Sailor2-20B and Meta-Llama-3-70B in these benchmarks. Notably, it scores 93.24% in M3Exam English, 72.97% in M3Exam Indonesian, and 70.03% in M3Exam Vietnamese, demonstrating elite global reasoning alongside strong regional linguistic nuance.

Good For

• Applications requiring efficient, high-performance LLMs in resource-constrained environments.
• Use cases demanding accurate and culturally appropriate responses in Southeast Asian languages, English, and Chinese.
• Tasks involving mathematical reasoning, translation, and complex instruction following.