Tencent Launches Hunyuan Hy3 LLM with Strong Enterprise Performance
- tech360.tv
- 5 hours ago
- 3 min read
Tencent has launched its Hunyuan Hy3 large language model. This model features a Mixture of Experts architecture. It includes 295 billion total parameters with 21 billion activated, supporting a 256K context length. The model reportedly achieves a 90 per cent task resolution rate on Tencent's WorkBuddy enterprise platform, according to Tencent Tech.
The Hy3 model employs a pragmatic approach. It prioritises real world performance over the largest possible parameter count. It demonstrates benchmark results that are competitive with models two to five times its size. This model is open sourced under the Apache 2.0 license, offering full commercial availability from its launch day zero on Hugging Face and Modelscope platforms.
Pricing for Hy3 is set at 1 yuan per million input tokens, 4 yuan per million output tokens, and 0.25 yuan per million tokens when using cache. This pricing positions it as a cost effective option within the Chinese market, particularly for enterprise agent applications. And the official version shows substantial improvements over its April preview.
Agent and code capabilities saw improvements of 20 to 30 per cent. Hallucination rates decreased by half. Gains were observed on SkillsBench, rising from 29.1 to 55.3, and on MathArena Apex, increasing from 12.8 to 38.7. On the ClawEval pass3 benchmark, Hy3 achieved a score of 68.5. This surpassed DeepSeek V4 Pro's 62.4 and Qwen 3.7 Max's 65.2. BrowseComp scores reached 84.2, nearly matching GPT 5.5's 84.4.
The model's effectiveness is attributed to its agent and tool orchestration capabilities. Internal evaluations on WorkBuddy indicated a rise in task resolution from 72 per cent to 90 per cent following the upgrade from the Hy3 preview to the official version. But the average time taken for tasks also reduced by 34 per cent. It demonstrated notably better performance in data processing, document handling, and research report analysis.
Ima's knowledge base QA and agent scenarios showed a system stability of 95.1 per cent. This included a significant reduction in blind retries and unnecessary operations. Real world case studies further illustrate Hy3's practical reliability, demonstrating its capability in complex operational tasks.
In one instance, the model constructed a consolidated cash flow model for an oil company, encompassing three regions, six reserve blocks, and 5,220 linked cells. Every number within this model operated as a live formula, automatically recalculating with assumption changes, thus requiring no hardcoded values. So, it demonstrated autonomous modelling capabilities.
Another application involved Hy3 generating complete Excel modelling from 101 SKUs of sales data. This output included 12 analysis tables and a 30 slide presentation deck featuring 20 charts, covering bar, pie, heat map, and bubble visualisations. These ranged from simple bar graphs to complex bubble charts.
Hy3 is now integrated into several Tencent products. These include WorkBuddy, CodeBuddy, Yuanbao, ima, Marvis, QQ Browser, Tencent News, WeGame, and Sogou Input. And nearly 50 more products are currently in the integration pipeline, expanding its operational footprint across the Big Tech organisation's portfolio.
Tencent has launched Hunyuan Hy3, a large language model with 295 billion parameters and 21 billion active parameters, achieving a 90 per cent task resolution rate on WorkBuddy.
The model prioritises real world performance, showing competitive benchmark results against larger models while maintaining aggressive pricing in yuan.
Hy3 demonstrates significant improvements in agent and code capabilities, with hallucination rates reduced by half compared to its preview.
Practical reliability is evidenced through real world applications, including complex financial modelling and comprehensive sales data analysis.
The model is integrated into numerous Tencent products and is scheduled for further deployment across nearly 50 additional offerings.
Source: Pandaily
