Morgan Stanley has released a research report noting that TENCENT (00700) launched and open-sourced its next-generation large language model, Hy3 Preview, on the 23rd. This release is positioned as the first step in rebuilding the Hunyuan (HY) foundation model. The model features a Mixture of Experts (MoE) architecture, integrating fast and slow thinking, with a total parameter count of 295B (21B activated parameters) and a context window of 256K. It demonstrates improvements in complex reasoning, instruction following, in-context learning, code generation, agent capabilities, and reasoning performance. The firm has assigned TENCENT an "Overweight" rating with a target price of HK$650. The report highlighted significant gains in code generation and agent capabilities for Hy3 Preview, achieving a score of 74.4% on the SWE-Bench Verified benchmark, compared to 53% for Hy2. This performance is roughly on par with GLM 4.7's score of 73.8%. TENCENT's comprehensive optimizations offer the best cost-performance ratio, with inference efficiency improved by up to 40% and overall costs substantially reduced. The model has already been integrated into TENCENT products such as Yuanbao, CodeBuddy, WorkBuddy, QQ, and ima. Competitive API pricing and customized token plans are available on Tencent Cloud, with a personal version starting as low as RMB 28 per month. Morgan Stanley indicated that TENCENT plans to use open-source feedback from developers and users to refine the official version. Subsequent steps will involve expanding training and reinforcement learning, deepening product ecosystem integration, and pursuing differentiated model capabilities to enhance real-world applicability.
Comments