Hy3 (295B A21B), a reasoning and agent model with great cost efficiency (github.com)

Tencent has launched Hy3, an advanced reasoning and agent model featuring 295 billion parameters, of which 21 billion are actively employed. This Mixture-of-Experts (MoE) model significantly enhances capabilities in various productivity tasks like coding, financial modeling, and design. The update follows extensive feedback and post-training improvements, demonstrating that Hy3 not only outperforms models of comparable size but also rivals larger flagship tools, achieving 2-5x efficiency in performance and utility. The significance of Hy3 lies in its major advancements in reliability and contextual understanding. It achieved a notable drop in hallucination rates from 12.5% to 5.4% and improved multi-turn intent tracking, decreasing error rates from 17.4% to 7.9%. The model's design includes an intricate architecture with 80 layers and 64 attention heads, allowing for a context length of 256K and streamlined tool interactions. As a public release under the Apache License 2.0, Hy3 is poised to offer developers an efficient, versatile solution for complex AI applications.