Skip to main content

Tencent's Hy3: Tiny Model, Giant Coding Skills, Matching DeepSeek-V4-Pro

In the world of AI language models, bigger isn't always better. That's the takeaway from SuperCLUE's latest programming evaluation, which pitted Tencent's Hy3 against some of the biggest names in the game—including DeepSeek-V4-Pro. And the results? Hy3, with its modest 21 billion activated parameters (out of 295 billion total), tied with DeepSeek-V4-Pro on coding ability, scoring 47.37. That's no small feat, especially when you consider that many of its competitors have several times more parameters.

Small Package, Big Punch

Hy3 uses a Mixture of Experts (MoE) architecture, which means it only activates a fraction of its total parameters for each task. This design choice pays off in spades. Not only does it keep costs down—completing a single programming question costs an average of just 0.43 yuan—but it also delivers speed. On average, Hy3 takes less than 400 seconds per question, making it suitable for real-time coding and online debugging.

Real-World Coding, Not Just Benchmarks

This evaluation wasn't your typical benchmark. It was designed to mimic real-world coding scenarios that Chinese programmers face daily. Each task required dozens of back-and-forth communications, gradually completing code analysis, modification, and verification. That's a lot closer to actual development and debugging than a simple one-shot test. The evaluation compared models across four dimensions: usage cost, runtime speed, communication rounds, and token consumption. This gives developers a practical sense of which model fits their budget and workflow.

Efficiency That Speaks Volumes

Hy3's communication efficiency is also impressive. It needed an average of over 40 conversation rounds to complete a task, without requiring repeated instruction adjustments to fix code. That means faster convergence and less hassle. Each task consumed only about 1.16 million tokens, which translates to lower cloud computing costs and less pressure on platform operations.

The Bottom Line

For programmers who rely on AI-assisted development, Hy3 offers a rare balance between performance and cost. It proves that optimized architectures and training strategies can compete with flagship-level opponents. And with Hy3 now fully open-sourced under the Apache 2.0 license, this cost-effectiveness battle is just getting started.

Key Points

  • Coding Ability: Hy3 scores 47.37, tying with DeepSeek-V4-Pro.
  • Cost: Only 0.43 yuan per question, making it affordable for frequent use.
  • Speed: Under 400 seconds per question on average.
  • Efficiency: Over 40 conversation rounds per task, with 1.16 million tokens consumed.
  • Open Source: Released under Apache 2.0 license.