Tencent Hy3 matches pricey models on code tests for zero cost

Tencent Hy3 recently exited preview and hit public platforms like OpenRouter entirely free of charge. In the past, free tier models struggled to even register on competitive LLM coding benchmarks, typically scoring only one or two points. Let's break down how this new contender performs across a rigorous five-project benchmark to see if it stands up to established, premium models.

The Core Tech Stack Divide

When we look at the results, the model shows a massive performance split depending on the technology stack. In a standard Laravel API generation test, Hy3 completed the prompt in under two minutes with a perfect score on three out of five runs. The failed attempts struggled with N+1 query optimization, a performance issue rather than broken syntax.

However, Filament—a less common admin panel framework—proved to be its Achilles' heel. Because eastern Chinese models rarely train heavily on smaller frameworks, Hy3 failed miserably here, scoring zero points overall. It proves a vital point: social media hype about a model being "good" or "bad" always depends on the specific codebases used for training.

Shining in React and Handling Edge Cases

Where the model truly shines is with mainstream web standards. Tested on a React and TypeScript component creation prompt with Playwright tests, Hy3 scored a four out of five. It generated clean code faster than the average of modern frontier models, taking just about one minute per run.

Even more surprising was the CSV importer challenge, which tests whether an LLM can anticipate complex edge cases without explicit prompting. Hy3 earned a 1.4 out of 5, matching the performance of Claude 3.5 Sonnet on the exact same project. Comparing a completely free model to Anthropic's expensive API tier reveals just how fast the performance gap is closing.

The Verdict on Tencent's New Challenger

With a total leaderboard score of 10.4, Tencent Hy3 sits near the bottom of the elite bracket. Yet, it managed to overtake Qwen 2.5 (referred to as Quen 3.7) and ran neck-and-neck with DeepSeek. For a model that costs absolutely nothing until its free tier expires on July 21st, it delivers highly respectable code. Just be sure to run automated tests to catch the occasional optimization error.