Tencent Hunyuan Open Source Mixture-of-Experts Model Hy3
Tencent Hunyuan officially launches the Hy3 large model. Based on improvements to the preview version's post-training data quality and RL compute scale, it shows significant progress in reasoning, agents, and long-context tasks. Its performance rivals flagship models with 2-5 times the parameter scale. Hallucination rate dropped from 12.5% to 5.4%, common sense error rate from 25.4% to 12.7%, and multi-turn question error rate from 17.4% to 7.9%.
By Xiao Jing
Edited by Xu Qing阳
On July 6, the official version of Tencent Hunyuan Hy3 was released. Hy3 is a model that fuses fast and slow thinking, adopts a MoE architecture, has a total of 295B parameters, 21B active parameters, and supports a 256K context length.
The Hy3 preview released on April 23 achieved a qualitative leap compared to Hy2 in complex reasoning, instruction following, context learning, code generation, and agent capabilities; Hy3 continues the clear and steep capability growth curve. By further increasing the post-training compute scale as well as data quality and diversity, it has leaped again across various tasks compared to the Hy3 preview, rivaling the performance of domestic and international large-scale flagship models for the first time with a smaller size.
Hy3 continues the practical and inclusive model positioning, priced at 1 CNY per million tokens for input, 4 CNY per million tokens for output, and only 0.25 CNY per million tokens for input cache hits.
Regarding open source, Hy3 adopts the most commercially friendly Apache 2.0 open source license, offering higher openness and freedom, allowing global developers to download and use it for commercial purposes for free. To further facilitate global developers, Hy3 will be launched on multiple overseas platforms successively, covering OpenRouter, Hermes, Kilo, Cline, OpenClaw, OpenCode, CherryStudio, etc., and simultaneously accessing the open source model community Huggingface and Modelscope platform on "day 0".
This set of key information defines Hy3's competitive coordinates, pursuing practicality rather than the largest parameter scale, but actual performance approaching that of flagship models with 2 to 5 times the parameters, while having absolute competitiveness in price.
Tencent Hunyuan Hy3 Release: Agent Capabilities and Product Experience Leap 01 Two Months, From Preview to Official Launch
The Hy3 preview released on April 23 was the first version after Hunyuan's reconstruction. After the preview version went online, it was truly called by multiple product lines within Tencent and used by global developers, repeatedly verified in real tasks. The post-training iteration over these two months focused on three directions: further increasing the post-training compute scale, improving post-training data quality and diversity, and performing fine-grained detection and training constraints for hallucination issues exposed during the preview stage.
The results are reflected in the benchmarks. In 12 horizontal comparisons, the two areas where Hy3 improved the most compared to the preview version are SkillsBench (from 29.1 to 55.3) and MathArena Apex (from 12.8 to 38.7). Agent and code core capabilities improved by 20%-30%, and the hallucination rate dropped by half.
In Agent and tool orchestration tests, Hy3 established a clear differentiated advantage.
ClawEval pass^3 scored 68.5, exceeding DeepSeek V4 Pro's 62.4 and Qwen 3.7 Max's 65.2; SkillsBench 55.3 also leads these two competitors; BrowseComp 84.2 is almost on par with GPT 5.5's 84.4. In code tests, SWE-bench Pro improved from 46.0 in the preview to 57.9, and NL2repo from 35.3 to 45.6, with large improvement margins, but still has a 3-5 point gap with DeepSeek V4 Pro and Qwen 3.7 Max.
Mathematical reasoning is currently the most obvious ceiling. On MathArena Apex, Hy3 scored 38.7, while GPT 5.5 is 85.4. In pure reasoning tasks, the structural gap between domestic models as a whole and GPT 5.5 still exists—Qwen 3.7 Max's 44.5 is also far below GPT 5.5. This is also a direction that domestic models as a whole need to catch up on.
02 Validation in Real Business
Hy3 has been integrated into WorkBuddy/CodeBuddy, Yuanbao, ima, Marvis, QQ Browser, Tencent News, WeGame, Tencent Lexiang, Sogou Input Method, WeChat Official Accounts, and nearly 50 other businesses are in the integration queue.
On WorkBuddy, the number of users who voluntarily chose Hy3 preview grew 6 times. After Hy3 went online, internal evaluations based on WorkBuddy office scenarios showed that compared to Hy3 preview, the task resolution rate jumped from 72% to 90%, and average time spent was shortened by 34%, with better performance in multiple细分 tasks such as data processing, document processing, and research report analysis.
ima evaluated Hy3 based on two core scenarios: online knowledge base Q&A and Agent. In Agent tasks, Hy3 performed excellently overall, with system stability as high as 95.1%, where tool orchestration capability was particularly outstanding. Invalid operations such as blind retries and stopping when not supposed to stop were significantly reduced, making complex office task planning more accurate and one-step. The knowledge base Q&A scenario was also significantly enhanced, with reasoning quality net improving by nearly 19%, thinking more systematically, information coverage more comprehensive, and the structural integrity and usability of long-text writing and plan generation significantly enhanced.
Yuanbao's conversational interaction scenarios also provided extremely valuable feedback for the model, such as the hallucination issues of Hy3 preview in long-text and AI search scenarios. Hunyuan, through deep cleaning of training data samples, fine-grained hallucination detection and training constraints, let the model learn to output stably under complex evidence.
In internal long-text and RAG evaluations based on real business scenarios, the hallucination rate of the Hy3 official version dropped by approximately 44% compared to the preview version. In online evaluations based on Yuanbao's real user logs, the common sense error rate of the Hy3 official version dropped significantly compared to the preview version, with the error rate reduced by 12.3% in deep reasoning (high) mode and by 8.5% in ultra-fast (no_think) mode.
After upgrading to Hy3, Yuanbao Agent advanced comprehensively in evaluations covering scenarios such as information query, data processing, document office, life decisions, and webpage creation. Tool orchestration is more stable, invalid retries and empty spinning are significantly reduced, and complex tasks can also be planned more accurate and in one step. In the Benchmark evaluation of Yuanbao Agent, Hy3 has surpassed a large number of excellent domestic models such as GLM 5.1 in the two Agent scenarios of comprehensive office and life services, approaching the capability performance of leading overseas models such as Claude Sonnet 4.6, sufficient to stably support real business links.
03 "Finished Product" Actual Test Cases
In WorkBuddy, using Hy3 to independently build a consolidated cash flow model for an oil and gas company—a master table linking 5220 cells across three regions and six reserve blocks, involving real modeling difficulties such as price switching year by year, NGL conversion by difference rate, zero production anomaly handling, and three-level subtotal summary. Hy3 did everything correctly, and met two key standards: zero hardcoding, every number is a live formula that can automatically recalculate with assumptions, indicating it is building the model autonomously rather than copying answers; the ultra-large model with thousands of lines across multiple tables maintained logical consistency throughout, without any single error causing chain pollution across thousands of cells.
For industries relying on complex models such as finance and consulting, this reliability directly determines whether the model can enter real workflows.
Author statement: This video was generated by AI
Another case covers the full link from scattered data to a reporting finished product. Inputting 101 SKUs of sales data into Hy3, requiring the output of Excel modeling analysis and a 30-page reporting PPT. The model constructed a complete analysis draft of 12 tables (including heat maps, layered analysis, native charts), while delivering a 30-page PPTX—20 charts covering bar charts, pie charts, heat maps, bubble charts, with each page also having an opinion sentence, and clear analysis logic.
Finally, from dozens of files with scattered information to a finished product that can be directly taken for reporting, the model autonomously completed the full process of analysis, modeling, visualization, and typesetting.
Author statement: This video was generated by AI
From a development scenario perspective, the real test of code capability is not just understanding algorithms; the model needs to schedule multiple technology stacks simultaneously in a project, autonomously plan the architecture, and end-to-end produce a runnable complete engineering project.
Let Hy3 generate a camera gesture interaction page: open the camera and wave to blow an image into tens of thousands of light points, then make a fist to gather them back into the original image.
In this one webpage, it needs to run through three independent technologies simultaneously and link them precisely: MediaPipe gesture tracking, Three.js tens of thousands of particle 3D rendering and glow shaders, and AI real-time source image generation. The initialization order, data flow, and rendering pipeline of the three technologies all need to be autonomously planned by the model at the architecture level; any error in any link will crash the whole page. Hy3 official version passed on the first try, gestures truly follow the hand, and particle overlay glow remains smooth.
Author statement: This video was generated by AI
Let Hy3 end-to-end generate the official website homepage of a fictional clean energy company. The first screen is a Three.js real-time rendered 3D fusion core, which disassembles as the camera advances with page scrolling. The whole site has seven or eight blocks, involving GLSL shaders, Bloom glow post-processing, and scroll-driven 3D camera. Nearly 2000 lines of code, autonomously completed by the model through planning, image generation, coding, and self-checking, with no blank screens or white screens throughout. The produced page can directly reach the commercial standard of a technology company's Landing Page.
Author statement: This video was generated by AI
04 90% of Agent Tasks Can Be Executed with Hy3
The conclusion of Tencent's Co-design testing with massive products is: at least 90% of Agent tasks in major scenarios can be successfully executed using Hy3. The coverage includes office and life automation scenarios such as data processing, document generation, research report analysis, information query, webpage creation, and life decisions.
The meaning of this number is that for the daily Agent needs of the vast majority of enterprises and individual users, Hy3 is already sufficient. It is a model pursuing practicality, "able to do 90% of the work well, while the cost is an order of magnitude lower than flagship models". For teams that truly want to integrate AI into business processes, being able to work and low cost may be the two highest priority options.
The path from preview version to official version is also worth noting: first enter real products, be used by real users, find optimization directions from real problems, and then bring the results back to the products. This "business feeds model" R&D link also completed a full cycle within two months.
This page provides an editorial summary based on publicly available information. It is not a republished article. Use the source link below for the original report.