Tencent Launches Hunyuan-3 The AI Model That Prioritizes Reality Over Benchmarks.

Tencent Unveils "Hunyuan-3" (Hy3): A New Era of Efficiency in Large Language Models

Tencent has officially announced the launch of its latest AI model family, Hunyuan-3 (Hy3). The name "Hunyuan" (混元) translates to "the origin of the universe," signaling a foundational shift in Tencent’s AI strategy. This new iteration features a architecture of 295B-A21B and supports a massive 256K context window, driven by a proprietary Reinforcement Learning (RL) training framework.

Prioritizing Real-World Performance Over Benchmarks

In a bold move, Tencent’s technical report revealed that the team has moved away from over-relying on traditional public benchmarks, citing that these metrics are increasingly prone to "gaming" or data contamination. Instead, Tencent developed rigorous internal testing suites based on real-time, up-to-date examinations.

The results place Hy3 on par with major competitors like GLM-5 and Kimi-K2.5. However, the standout feature of Hy3 is its efficiency; it achieves these performance levels while being significantly smaller than its rivals. Tencent describes Hy3 as a foundational milestone for its future AI evolution.

Pricing and Availability

Hunyuan-3 is now integrated into Tencent’s ecosystem and is available via API. The pricing is highly competitive, starting at 1.2 RMB (Input) / 4 RMB (Output) per million tokens.

The architecture specification of 295B-A21B suggests that Hy3 may utilize a Mixture of Experts (MoE) system. The number 21B represents the number of "active" parameters used at a time, explaining its energy efficiency and speed, even with a large model size of 295B.

Tencent's criticism of public benchmarks as "easily manipulated" is bold and straightforward. This reflects developers' increasing focus on "human evaluation" rather than inflated scores that lack real-world usability (Goodhart's Law).

The very low API pricing reflects Tencent's confidence in its own cloud infrastructure. Given its extensive ecosystem (WeChat, video games, FinTech), having intelligent yet resource-efficient models allows for significantly lower costs in embedding AI across all its services compared to competitors.

 

Google Gemini Can Now Generate and Download Files Directly from Your Chat. 

 

Source: Tencent Hy 

💬 AI Content Assistant

Ask me anything about this article. No data is stored for your question.

Tencent Unveils "Hunyuan-3" (Hy3): A New Era of Efficiency in Large Language Models

Tencent has officially announced the launch of its latest AI model family, Hunyuan-3 (Hy3). The name "Hunyuan" (混元) translates to "the origin of the universe," signaling a foundational shift in Tencent’s AI strategy. This new iteration features a architecture of 295B-A21B and supports a massive 256K context window, driven by a proprietary Reinforcement Learning (RL) training framework.

Prioritizing Real-World Performance Over Benchmarks

In a bold move, Tencent’s technical report revealed that the team has moved away from over-relying on traditional public benchmarks, citing that these metrics are increasingly prone to "gaming" or data contamination. Instead, Tencent developed rigorous internal testing suites based on real-time, up-to-date examinations.

The results place Hy3 on par with major competitors like GLM-5 and Kimi-K2.5. However, the standout feature of Hy3 is its efficiency; it achieves these performance levels while being significantly smaller than its rivals. Tencent describes Hy3 as a foundational milestone for its future AI evolution.

Pricing and Availability

Hunyuan-3 is now integrated into Tencent’s ecosystem and is available via API. The pricing is highly competitive, starting at 1.2 RMB (Input) / 4 RMB (Output) per million tokens.

The architecture specification of 295B-A21B suggests that Hy3 may utilize a Mixture of Experts (MoE) system. The number 21B represents the number of "active" parameters used at a time, explaining its energy efficiency and speed, even with a large model size of 295B.

Tencent's criticism of public benchmarks as "easily manipulated" is bold and straightforward. This reflects developers' increasing focus on "human evaluation" rather than inflated scores that lack real-world usability (Goodhart's Law).

The very low API pricing reflects Tencent's confidence in its own cloud infrastructure. Given its extensive ecosystem (WeChat, video games, FinTech), having intelligent yet resource-efficient models allows for significantly lower costs in embedding AI across all its services compared to competitors.

 

Google Gemini Can Now Generate and Download Files Directly from Your Chat. 

 

Source: Tencent Hy