Tencent's New Hy3 AI Model: An Efficient Chinese Large Language Model

Tencent quietly unveiled its most advanced AI model to date this Thursday, backed by impressive benchmark results. The Hy3 Preview, the first release following a comprehensive infrastructure overhaul, is now open-source on GitHub, Hugging Face, and ModelScope.

Additionally, the model is accessible via paid tiers on Tencent Cloud’s official platform.

Boasting 295 billion total parameters—a proxy for knowledge capacity—Hy3 activates only 21 billion at any given moment. This efficiency stems from its Mixture-of-Experts (MoE) architecture, which dynamically routes queries to specialized sub-networks rather than processing everything simultaneously. The approach drastically cuts computational overhead and costs while maintaining output parity. It also supports a massive 256,000-token context window, enabling the processing of full-length novels within a single prompt.

Designed to harmonize three critical dimensions—broad capability, transparent evaluation, and cost-efficiency—the model marks a strategic pivot from Tencent’s previous 400-billion-parameter Hy2. The company explicitly notes that 295 billion represents the optimal threshold where reasoning capabilities mature fully, and further parameter scaling yields diminishing returns.

A leaner parameter count does not equate to inferior performance. In fact, meticulously trained, smaller-scale models frequently outperform bulkier, generalized counterparts.

Coding capabilities show remarkable advancement. On SWE-bench Verified, which evaluates real-world bug resolution in production repositories, Hy2 scored 53.0%, while Hy3 Preview jumps to 74.4%. This 40% generational leap places it near Claude Opus 4.6 (80.8%) and surpasses both GLM-5 (77.8%) and Kimi-K2.5 (76.8%). Similarly, Terminal-Bench 2.0, measuring autonomous CLI task execution, surged from 23.2% to 54.4%.

For developers building AI agents, this model presents a compelling option. Agents require intricate orchestration of memory, tool usage, and skill execution, where minor oversights can derail workflows. As agent-centric development becomes an industry focal point, robust agent compatibility is crucial. Consequently, Hy3 Preview is immediately integrated with Openclaw.

Autonomous search and browsing capabilities have also seen substantial upgrades. On BrowseComp, which tracks complex web research tasks, Hy3 Preview achieves 67.1%, up from Hy2’s 28.7%. On WideSearch, it scores 70.2%, outperforming GLM-5 and Kimi-K2.5, though trailing Claude Opus 4.6’s 77.2%.

In mathematical and scientific reasoning, the model outperformed all domestic rivals on the Tsinghua University Ph.D. qualifying exam (Spring 2026), averaging 88.4 across three runs. Tencent emphasizes real-world assessments over curated datasets to prevent benchmark manipulation. It also secured an 87.8 on the 2025 CHSBO (China High School Biology Olympiad), the highest score among Chinese models in this category.

Training commenced in late January 2026, culminating in Thursday’s release—a sub-three-month turnaround from cold start to open-source launch. Tencent attributes this accelerated timeline to a foundational infrastructure revamp led by Chief AI Scientist Shunyu Yao in February, which completely rebuilt the pre-training and reinforcement learning pipelines.

This rapid development cycle contrasts sharply with the industry landscape a year ago, when DeepSeek’s R1 disrupted the market through unprecedented cost-efficiency.

While Hy3 still trails behind flagship offerings from OpenAI and Google DeepMind, its performance-to-size ratio is highly competitive. Agent benchmarks position it in the "sweet spot" with ~295B parameters, outpacing DeepSeek-V3.2 (600B+) and matching Kimi-K2.5 (1T+) at a fraction of the computational expense.

The Hunyuan series is already deployed across Tencent’s ecosystem, including Yuanbao, CodeBuddy, WorkBuddy, QQ, and Tencent Docs. On CodeBuddy and WorkBuddy, time-to-first-token latency dropped by 54%, and end-to-end generation time decreased by 47%, successfully handling agent workflows up to 495 steps. Tencent Cloud offers API access at approximately $0.18 per million input tokens and $0.59 per million output tokens, with personal tier subscriptions starting around $4.10 monthly.