Tencent’s latest artificial intelligence model, the Hy3 Preview (also known as Hunyuan 3.0 in the domestic market), has generated a wave of positive feedback from the developer community following its open-source release. The highly capable Mixture-of-Experts (MoE) system represents a significant leap forward for the Chinese technology giant, though an exclusive report from The Information suggests that the model’s impressive performance was achieved, in part, through the use of Anthropic’s Claude for fine-tuning.
The launch of the Hy3 Preview on April 24, 2026, marked a critical milestone for Tencent as it seeks to close the gap with domestic rivals like ByteDance and Alibaba, as well as global leaders like OpenAI. The model’s reception is particularly noteworthy given the intense scrutiny surrounding its development under the leadership of Yao Shunyu, a former OpenAI researcher who recently joined Tencent as Chief AI Scientist.
A Rebuilt Infrastructure and Impressive Specs
According to Tencent’s official press release, the Hy3 Preview is the most intelligent model in the Hunyuan series to date. The company undertook a massive engineering effort, completely rebuilding its pre-training and reinforcement learning infrastructure from scratch starting in February 2026. This overhaul was designed to make the models more practical for real-world use, focusing on complex reasoning, long-context understanding, and agentic capabilities.
The technical specifications of the Hy3 Preview are formidable. It is a massive MoE model with 295 billion parameters, of which 21 billion are activated during inference. It supports a massive 256K token context window, allowing it to process and analyze extensive documents and codebases. Tencent claims the new architecture delivers a 40 percent improvement in inference efficiency compared to its predecessor, achieving optimal intelligence density at a comparable cost.
The model has already been deeply integrated across Tencent’s core product ecosystem, including the Yuanbao AI chatbot, QQ, Tencent Docs, and its coding assistants CodeBuddy and WorkBuddy. In these practical applications, the Hy3 Preview has demonstrated substantial performance gains. For instance, in CodeBuddy, the Time To First Token (TTFT) was lowered by 54 percent, and the end-to-end response time decreased by 47 percent, while maintaining a success rate exceeding 99.99 percent. Furthermore, the model has proven capable of reliably powering complex agent workflows of up to 495 steps.
(Related: Tencent Launches Hunyuan Hy3 in 90 Days and Hires Former OpenAI Researcher to Lead It)
The Anthropic Connection
While Tencent’s official communications highlight the company’s internal engineering achievements and the leadership of Yao Shunyu, The Information’s report adds a crucial layer of context to the model’s success. According to the publication, Tencent employees utilized Anthropic’s Claude, one of the most advanced Western AI models, to fine-tune the Hy3 Preview.
This practice, often referred to as “AI distillation” or using a “teacher model” to train a “student model,” is increasingly common in the rapidly evolving AI industry. By leveraging the outputs and reasoning capabilities of a highly capable model like Claude, developers can significantly accelerate the training process and improve the performance of their own systems, often at a fraction of the cost of training from scratch.
The use of Anthropic’s technology by a major Chinese tech giant underscores the porous nature of the global AI ecosystem, despite escalating geopolitical tensions and US export controls. It also highlights the ongoing debate within the AI community regarding the ethics and long-term viability of relying on competitors’ models for training data. While distillation can provide a rapid performance boost, critics argue that it may ultimately limit a model’s ability to surpass its “teacher” capabilities.
Open Source Strategy and Market Positioning
Tencent’s decision to open-source the Hy3 Preview is a strategic move designed to accelerate adoption and foster a robust developer ecosystem around its technology. The model is now available on major platforms including GitHub, Hugging Face, ModelScope, and GitCode, and it supports mainstream inference frameworks such as vLLM and SGLang.
To further incentivize developers, Tencent has made the Hy3 Preview API available on OpenRouter for free for a limited two-week period. On its own Tencent Cloud TokenHub platform, the company is offering highly competitive pricing, with input costs starting at approximately $0.18 per million tokens and cached input costs at just $0.06 per million tokens.
This aggressive pricing and open-source strategy position Tencent to compete directly with other leading Chinese AI developers, such as DeepSeek, which recently slashed prices for its V4-Pro model. By making its most advanced model widely accessible and highly affordable, Tencent aims to capture a larger share of the enterprise and developer market, solidifying its position as a dominant force in China’s generative AI landscape. The positive early reviews suggest this strategy is gaining traction, though the long-term success of the Hy3 Preview will depend on its ability to maintain its competitive edge in a fiercely contested market.