On April 23, Tencent released and open-sourced its Hunyuan Hy3 preview language model. It's a mixture-of-experts model that blends fast and slow thinking, with 295 billion total parameters, 21 billion active parameters, and a maximum context length of 256K tokens. This is the first model trained after Hunyuan's infrastructure rebuild, and it's the smartest model Hunyuan has ever produced, with significant improvements in complex reasoning, instruction following, in-context learning, coding, agent capabilities, and inference performance.
In February 2026, Tencent rebuilt its pre-training and reinforcement learning infrastructure, along with three principles for pursuing practicality:
1. Systematic capabilities: The company doesn't advocate for "specialization" — even a single code agent application involves deep coordination of reasoning, long-form text, instruction following, dialogue, code, tools, and more.
2. Authentic evaluation: It actively avoids public leaderboards that are easy to game, instead using self-built questions, the latest exams, human evaluation, and product crowd-testing to assess and improve the model's "real combat effectiveness."
3. Cost-performance pursuit: Practicality requires commercial viability. Deep coordination between model architecture and inference framework design significantly reduces task costs, making intelligence affordable and usable.
Hy3 preview can be seen as the beginning of Hunyuan's rapid exploration of practical large models that solve real-world problems.
Tencent's chief AI scientist, Yao Shunyu, said, "Hy3 preview is the first step in rebuilding the Hunyuan large model. We hope that through this open-source release, we can get real feedback from the open-source community and users to help us improve the practicality of the official Hy3 version." At the same time, the company is continuing to scale up pre-training and reinforcement learning to raise the model's intelligence ceiling, and through deep co-design with many Tencent products, it is continuously improving the model's comprehensive performance in real scenarios and beginning to explore distinctive model capabilities.
Hy3 preview is now available on Tencent Cloud, Yuanbao, ima, CodeBuddy, WorkBuddy, QQ, QQ Browser, Tencent Docs, and Tencent LeShare, with more products like WeChat Official Accounts, Game for Peace, Tencent News, Tencent Self-Selected Stocks, Tencent Customer Service, and WeRead rolling out. It also supports popular open-source agent products like OpenClaw, OpenCode, and KiloCode, and is listed on Tencent Cloud's large model service platform TokenHub.
Hy3 preview focuses on comprehensive practicality, with significantly improved agent capabilities
Multiple evaluation results show that Hy3 preview's capabilities are comprehensively improved.
1. Excellent in-context learning and instruction following
In various real-world production and life scenarios, understanding messy, lengthy context and following complex, changing rules is the primary challenge for models. Inspired by Tencent's business scenarios, Hunyuan proposed CL-bench and CL-bench-Life to innovatively evaluate the model's in-context learning ability, and Hy3 preview significantly improved its in-context learning and instruction following.
2. Outstanding complex reasoning, highest score on Tsinghua math PhD qualifying exam
Complex reasoning is the foundation for models to solve various problems. Hy3 preview performs well on high-difficulty STEM reasoning tasks like FrontierScience-Olympiad and IMOAnswerBench, and achieved excellent results on the latest Tsinghua University Qiuzhen College math PhD qualifying exam (Spring 2026) and the National High School Biology League (CHSBO 2025), demonstrating generalizable strong reasoning.
3. Code and agents see the most significant improvement, showing high cost-performance
Code and agents are the areas where Hy3 preview has improved the most. Thanks to the rebuilt pre-training and reinforcement learning frameworks and the increased scale of reinforcement learning tasks, Hunyuan quickly achieved competitive results on mainstream code agent benchmarks like SWE-Bench Verified and Terminal-Bench 2.0, as well as search agent benchmarks like BrowseComp and WideSearch.
In the digital world, code focuses on the model's execution ability in a development environment, while search focuses on retrieval, filtering, and integration in open information spaces. Together, they determine whether the model is truly usable in complex agent scenarios (e.g., OpenClaw). Hy3 preview performs well on evaluations like ClawEval and WildClawBench, indicating that its agent capabilities are steadily moving toward comprehensiveness and practicality.
Beyond public benchmarks, Hunyuan has also built multiple internal evaluation sets to assess the model's performance in real development scenarios. Results show that Hy3 preview is highly competitive on the backend engineering task set Hy-Backend, the user-interaction-focused Hy-Vibe Bench, and the high-difficulty software engineering development task set Hy-SWE Max.
Comparing the size and overall agent performance of various open-source models, Hy3 preview demonstrates high cost-performance.
Tencent's core businesses are fully integrated, with clear positive returns across multiple AI products
Before its official launch, Hy3 preview was tested on Tencent's major AI products and showed clear positive returns.
On the Yuanbao side, Hunyuan and Yuanbao conducted deep co-design. On one hand, they specifically improved the model's performance on hard metrics like intent understanding accuracy, text creation quality, and deep search. On the other hand, they fine-tuned writing style, emotional intelligence, content organization, and professionalism. The deep synergy between model and product brings users a smarter, more "human-like" interaction experience.
In two scenarios — ima knowledge base Q&A and general Q&A — test results show that Hy3 preview excels at handling long texts, especially retrieval tasks, with good accuracy, coverage, and comprehensiveness.
On CodeBuddy and WorkBuddy, Hy3 preview reduces first-token latency by 54%, end-to-end time by 47%, and increases success rate to 99.99%+. In real user environments, Hy3 preview has stably driven complex agent workflows up to 495 steps, covering diverse office scenarios like document processing, data analysis, knowledge retrieval, and MCP toolchain orchestration.
In specialized evaluations for WeChat Official Account AI avatars and AI customer service, Hy3 preview shows a more comprehensive capability upgrade compared to Hy2. The new model is more mature in user intent understanding, complex context handling, and knowledge organization. When faced with vague questions, short follow-ups, and multi-turn conversations, it can more accurately grasp user needs and output clearer, more stable responses. When generating answers combined with knowledge bases, user memory, and context, it better fits the role of AI avatar or customer service, with significantly reduced over-imagination, subjective projection, and emotional expression, making the overall interaction closer to the goal of "trustworthy, natural, and efficient" responses.
In the Game for Peace AI NPC evaluation, the Game for Peace team integrated and evaluated Hy3 preview immediately after its launch, and the overall performance was impressive. In out-of-game role-playing scenarios, Hy3 Preview not only accurately understands character settings but also outputs highly relevant, value-added content for open-ended questions, delivering a more realistic, natural, and immersive dialogue experience. In complex in-game combat scenarios, the model's response rhythm is close to real player chat, showing excellent stability and impressive human-like role-playing ability.
In the Tencent Docs AI PPT scenario, compared to the previous version (Hy2), it achieved significant progress: generation success rate increased by 20%, evaluation score improved by 10%, and generation time decreased by 20%. Overall, the new model performs excellently in evaluation scenarios, showing strong performance in template selection, color matching, outline generation, and content supplementation — no hallucinations, on-topic, and visually appealing.
In the QQ AI assistant Xiao Q product evaluation, compared to the previous version, it significantly optimized long-text first-byte latency, overall response speed, and streaming output efficiency. In core capabilities, math reasoning improved notably, and multi-scenario instruction following and generalization were further enhanced. Tool-calling reasoning and multi-turn coreference resolution became more stable and efficient, achieving outstanding results in the OpenClaw official PinchBench QQ agent scenario test, with a clear leap in overall experience.
Inference efficiency improved by 40%, best intelligence density at the same cost
Thanks to deep coordination between model and inference framework, as well as comprehensive optimization in inference framework, operator performance, and quantization algorithms, overall inference efficiency has improved by 40%, and the cost of Hy3 preview has dropped significantly compared to the previous generation.
On Tencent Cloud's large model service platform TokenHub, Hy3 preview's input price is as low as 1.2 yuan per million tokens, input cache hit price is 0.4 yuan per million tokens, and output price is as low as 4 yuan per million tokens. Tencent Cloud has also partnered with Hunyuan to launch a customized Hy3 preview Token Plan package, with a personal edition starting at 28 yuan per month, offering a more cost-effective option for agent development and building "lobster" applications.