Tencent’s Apache‑Licensed Hy3 Takes on GLM‑5.2 with Half the Size — Wins in Search, Loses in Code

Hy3’s Apache 2.0 Release Redefines Open‑Model Access

Tencent has lifted the legal veil that hampered many Chinese AI models by issuing the full‑scale Hy3 under the permissive Apache 2.0 license. The 295‑billion‑parameter Mixture‑of‑Experts system activates only 21 billion parameters per token, yet it matches or exceeds larger rivals on several enterprise metrics.

Performance vs. GLM‑5.2

In a blind human evaluation involving 270 experts, Hy3 posted 2.67 against GLM‑5.1’s 2.51. The newer GLM‑5.2, however, still leads the coding suite, scoring 84.2 on SWE‑bench Verified versus Hy3’s 78.0. The gap aligns with the size differential: GLM‑5.2 runs on roughly 744 billion total parameters with about 40 billion active per token, while Hy3’s footprint is under 300 GB in FP8.

“Hy3’s real advantage lies in reliability, not raw code generation,” a Tencent AI lead told Reuters.

Enterprise‑focused metrics receive the spotlight. Hallucination rates fell from ↑ 7.1% (12.5%→5.4%) and commonsense errors halved, a result of stricter data cleaning and a “answer‑when‑grounded” policy. Multi‑turn dialogue quality jumped from 42.9% to 75.1% on the MRCR benchmark.

Deployment economics also tilt in Hy3’s favor. The model fits comfortably on Nvidia’s H20‑3e GPUs—hardware cleared for export to China—requiring a single 8‑GPU node versus the larger cluster needed for GLM‑5.2. This translates to lower capital outlay and easier scaling for global firms.

While independent verification is pending, the open‑source community has already flagged the license shift as the story’s headline. With no regional restrictions, Hy3 becomes a viable contender for any enterprise, whether navigating post‑pandemic digital transformation or building agent‑centric applications.