Tencent says Hy4 scores strongly on work tasks, with fresh benchmarks to test

Tencent says its Hy4 model performed well in workplace, financial and scientific evaluations, adding useful detail to the debate over what the new model can actually do. The reported results are company assessments, not independently established rankings, but they give readers more to examine than a claim to be among the best.

Watch Desk analysis

What happened

Hy4 adds visual understanding and computer-use capabilities, including work with professional software such as Blender and CAD, according to Tencent information reported by Lavx.hu. The report also describes integration with WeCom, where the model can draw on chats, email, documents, calendars and meetings for tasks such as tracking projects and deadlines.

Lavx.hu says Hy4 scored an average 83.18 with a 96.15% pass rate across WorkBuddy evaluations, ranking first in nine of twelve scenarios in China. It also reports a blind assessment involving 213 financial professionals, in which Hy4 ranked first among the open-source models tested. These are results attributed to Tencent, not independent confirmation that Hy4 is best across the market.

The report says Tencent has also released two models developed with Hy4: HyCell for single-cell representations and HyPDE for physics fields. Tencent describes its safety work as focusing on defensive cybersecurity and biological risks; the reported assessments do not establish how those safeguards perform in wider use.

Why it matters

Benchmarks can help distinguish a capable model from a glossy launch, but scores are only as informative as the tasks, comparison set and evaluation method behind them. The financial assessment and workplace results offer specific claims to scrutinise, while the WeCom integration points to a practical shift: an assistant connected to workplace records could help track work across several tools.

Our read

There is substance here, especially in the reported task results and the two research models. But Tencent’s own scores are a starting point for comparison, not the finish line. The useful question is whether independent users can reproduce the gains on tasks that matter outside Tencent’s own products.

What to watch

• Whether Tencent publishes evaluation methods and task-level results.
• How Hy4 compares with other models on the same financial and workplace tests.
• Whether the WeCom integrations and HyCell and HyPDE models see practical use.

Discussion spark: Should a model maker’s own benchmark results count as meaningful evidence of progress, or should they carry little weight until independent evaluators reproduce them?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.