Data Machina #247

Data Machina · 2 years ago

Open source mixture-of-experts models from AI21Labs, Alibaba, MetaAI, Databricks, and xAI are achieving near state-of-the-art performance comparable to closed models from OpenAI and Google. Databricks' DBRX uses 132B total parameters with 36B active per input, Alibaba's Qwen1.5-MoE-A2.7B reduces training costs by 75% while matching 7B model performance, and xAI's Grok-1.5 features a 128K context window. These efficient open MoE architectures allow researchers and developers to deploy competitive alternatives to proprietary large language models with reduced computational overhead.