On Thursday, Tencent unveiled its latest and most efficient AI model yet, dubbed Hy3 preview. This release marks the company’s first model following a comprehensive infrastructure overhaul and is now accessible via GitHub, Hugging Face, and ModelScope, with paid access on Tencent Cloud.
Hy3 preview boasts 295 billion total parameters but only activates 21 billion at any given time, leveraging a Mixture-of-Experts architecture. This design routes queries to specialized sub-networks, optimizing compute efficiency while maintaining quality output. It supports up to 256,000 tokens of context, enough for a full-length novel.
Tencent aimed to balance capability breadth, honest evaluation, and cost-efficiency with this model. Previously, their flagship Hy2 contained over 400 billion parameters; however, Tencent has since determined that 295 billion is the optimal parameter count where reasoning matures without incurring prohibitive costs.
Despite fewer parameters, Hy3 preview outperforms larger models by focusing on quality training. In coding tests like SWE-bench Verified—which evaluates a model’s ability to fix real bugs from GitHub repositories—Hy3 preview scored 74.4%, a significant improvement over Hy2’s 53.0%. This places it ahead of GLM-5 and Kimi-K2.5, though slightly behind Claude Opus 4.6.
On Terminal-Bench 2.0, which assesses task execution in command-line environments, the model’s score rose from 23.2% to 54.4%, marking another substantial advancement.
Hy3 preview is particularly suited for AI developers building agents that require complex instructions involving memories and tools. Its agentic capabilities have made it immediately available on Openclaw. The model also excelled in search tasks, scoring 67.1% on BrowseComp and 70.2% on WideSearch, surpassing several competitors.
In reasoning tests, Hy3 preview topped Chinese models on Tsinghua University’s math PhD qualifying exam (Spring 2026) with an average score of 88.4 across three runs. It also led in China’s national high school biology olympiad with a score of 87.8.
Training began in late January 2026, and it was launched within three months, thanks to an infrastructure overhaul by chief AI scientist Yao Shunyu. This rapid development contrasts with previous Chinese approaches, such as DeepSeek’s R1, which emphasized cost-efficiency.
While Hy3 still lags behind OpenAI and Google DeepMind in some areas, its size-to-performance ratio is noteworthy. The agent benchmark composite places it in the