Just now, Yao Shunyu led the team to “rebuild” Hunyuan, and the first large model has gone live.
Just now, Tencent Hy3 preview has arrived.
This is the first version of Hy3 released after Tencent’s Hunyuan team restarted from architecture and infrastructure. The initial batch of models is relatively small in size, positioned toward practicality.
It’s also worth noting that Hy3 preview is the first major result after Yao Shunyu returned to China and joined Tencent. It follows his idea of the “second half of AI,” where the Hy3 model is refined and improved within Tencent’s real business and complex scenarios, focusing on effectiveness and practicality in real-world use cases.
Tencent stated that the new generation Hy3 model enhances capabilities in chatting, coding, agents, mathematical reasoning, instruction following, and context understanding.
Hy3 Deployment Across Tencent Ecosystem
At present, Hy3 preview has already launched across Tencent Cloud, Yuanbao, ima, CodeBuddy, WorkBuddy, QQ, QQ Browser, Tencent Docs, and Tencent Lexiang. It is also being rolled out gradually across major products such as WeChat Official Accounts, Peacekeeper Elite, Tencent News, Tencent Self-selected Stocks, Tencent Customer Service, and WeChat Reading.
In addition, Hy3 preview supports integration with popular open-source agent tools such as OpenClaw, OpenCode, and KiloCode, and has been listed on Tencent Cloud’s large model service platform, TokenHub.
Before the May Day holiday, global AI teams have been making moves one after another. We’ve already seen Qwen 3.6 Max Preview, Moonshot AI’s Kimi 2.6, followed closely by Xiaomi’s MiMo-V2.5-Pro.
So how capable is Tencent Hy3 preview as a foundational large model? Next, we tested Hy3 hands-on.
Hy3 Agent Capability Stress Test: Did It Handle the “Lobster Task”?
Hy3 Multi-step Reasoning and Research
Yao Shunyu is the proposer of the ReAct framework (core logic for agents), so Hy3 preview’s improvements naturally include coding and agent capabilities, aligning with the technical trends and market demand of the agent-driven second half.
In Tencent’s AI office assistant WorkBuddy, we can ask Hy3 to perform coding, in-depth research, product management, data analysis, and more.
For example, we asked Hy3 preview to investigate rumors about DeepSeek’s financing, requiring it to compare at least five authoritative sources with different backgrounds, list known facts and logical conflicts, and provide confidence scores.
Hy3 was able to autonomously initiate multi-step searches, complete long-chain reasoning, systematically sort out contradictions among sources, and finally present an objective and neutral investigation report—all without human intervention.
Hy3 Data Analysis and Visualization Tasks
We further asked Hy3 preview to fetch data from the United Nations Population Division and complete a visualization analysis on “global population structure changes.”
This is a composite task involving data acquisition, cleaning, analysis, and visualization. Hy3 handled it relatively smoothly, ultimately producing intuitive charts and analytical text.
Hy3 Coding Ability in Practice
In coding capability testing, reflecting modern vibe coding practices, we asked Hy3 preview to generate a web-based “match-three” game.
The final result had polished visuals, complete logic, and ran normally. The overall quality from Hy3 exceeded expectations.
Hy3 Conversation and Creativity: Are the Fundamentals Solid?
We can see that Tencent Hunyuan, like other models, distinguishes between modes: fast thinking for quicker answers, and deep thinking for more thorough reasoning and higher-quality responses. Here, we used deep thinking throughout with Hy3 preview.
Hy3 Daily Chat and Emotional Understanding
This upgrade emphasizes practicality, so we first tried casual conversation with Hy3.
When we complained to Hy3 preview about feeling less sharp recently, it patiently suggested it might be due to lack of sleep, work pressure, or too much short-video scrolling, and gave three small tips.
Continuing the topic of lacking inspiration for writing, Hy3 naturally connected to the conversation context, adjusted its tone and depth based on our emotional state, and provided targeted creative suggestions.
It can also deliver strong emotional value, even complimenting in varied ways.
Hy3 Common Sense Reasoning
Previously, Zhihu hosted a discussion called “AI, take this challenge,” which included tricky questions where AI often fails. One question was: “I only found out this year that my biological parents didn’t invite me to their wedding. I feel upset. What should I do?”
Many large models got confused, ignoring the basic logic that the child wasn’t born yet at the time of the parents’ wedding. Hy3 preview quickly caught this point, guided the user to sort out emotions, and showed strong common-sense reasoning and empathy.
Hy3 Creative Writing and Style Mimicry
Hy3 Social Media Copywriting
Recently, a stunning image of NASA astronauts looking back at Earth through the Orion spacecraft window went viral.
We asked Hy3 preview to generate five social media captions for the image. It first analyzed the atmosphere, chose emotional angles like loneliness and awe, reverence for Earth, and the smallness and greatness of humanity, then produced captions in different styles—some literary, some philosophical. Any one of them from Hy3 could be posted directly.
Hy3 Writing in a Specific Style
For style imitation, we asked Hy3 to write a short story in the tone of O. Henry.
Hy3 Search and Information Reliability
For search capability, we asked Hy3 preview to investigate why Meta forcibly collects mouse and keyboard input. It quickly cited authoritative sources and provided a clear, evidence-based answer.
Whether checking news, policies, or verifying specific information, Hy3’s overall performance was relatively reliable.
Hy3 and the Reconstruction of the “Second Half of AI”
According to reports, Hy3 preview is a MoE (Mixture of Experts) language model that integrates fast and slow thinking. The Hy3 model has a total of 295B parameters, with 21B activated parameters, supports a context length of 256K, and balances practicality with cost efficiency.
In this new generation Hy3 model, the Hunyuan team’s main work focused on foundational reconstruction, making many aspects more solid—especially pretraining and reinforcement learning infrastructure, which have been completely rebuilt. Instead of pursuing minor innovations in attention mechanisms or base architecture, the Hy3 architecture follows a mature MoE route and invests heavily in strengthening the engineering infrastructure.
This means Hy3 preview’s stability, data throughput efficiency, and reinforcement learning pipeline yield may have reached an unprecedented industrial-grade level.
Hy3 Evaluation Over Training
During training, the Hy3 model emphasizes model evaluation and strengthens research on self-built benchmarks. This aligns with Yao Shunyu’s previously shared idea: evaluation is more important than training (Evaluation > Training).
In “The Second Half of Large Models,” Yao Shunyu pointed out that the current large model “recipe” (pretraining + reinforcement learning + scaling compute) is already highly mature, capable of generalization and solving complex problems. The key question in the second half becomes: “What should we train AI like Hy3 to do?”
Because the general model recipe is already extremely powerful, spending massive effort on fine-tuning may only yield about a 5% improvement. Therefore, evaluation becomes more important than training. The industry needs to reconstruct evaluation systems, designing new tasks and paradigms closer to real-world scenarios, rather than simply making harder test questions.
Hy3 From Technology to Real-World Problem Solving
To survive and grow in the second half of AI, practitioners must shift their mindset and adopt a “product manager” perspective. This means deeply thinking about: who should AI like Hy3 serve, what real problems it should solve, and how we objectively measure whether it solves them well.
In this regard, Tencent has some of the most complex business scenarios in China and even globally—covering WeChat, gaming, advertising, and cloud services. Its self-built evaluation environment is naturally highly aligned with real business flows, challenges, and pain points.
The release of Hy3 preview may already represent a productivity tool within Tencent’s ecosystem that can solve real-world problems.
Hy3 Is Just the Starting Point
Hy3 preview began training at the end of January 2026, and it took less than three months from training to launch. This marks a shift for the Hy3 model—from “reading thousands of books” to “traveling thousands of miles,” starting to tackle real-world problems.
Hy3 preview is just a starting point. In the future, the Hunyuan team hopes to further improve Hy3 capabilities through collaboration with developers and users, allowing the Hy3 ecosystem to continue evolving in real scenarios and tasks.
