Tencent open-sources Hy4 model with finance and agentic focus

Tencent's Hy4 targets complex financial workflows and agentic tasks, with a hallucination rate the company says it cut to 7.2% in internal tests.

By The Datatech Times Editorial Team AI level 2 of 5
AI level 2 of 5: drafted by our AI editorial assistant from the company's announcement; fact-checked, edited and signed off by Mark Walker, Editorial Director. What the levels mean
A modern control room features a large wall of multiple screens displaying abstract digital patterns, a control desk with several monitors, buttons, and joysticks, an office chair, and two potted plants, all brightly lit by large windows.

Tencent has open-sourced Hy4, the latest generation of its Hunyuan large language model series, adding visual understanding, computer-use capabilities and deep integration with the company's enterprise productivity suite. The release, dated 4 November 2026, positions Hy4 as a broad-purpose model while placing particular emphasis on financial analysis workflows, a differentiator that the company backed with a structured external benchmark.

Tencent says model weights and system cards are due to be published shortly after the main announcement. The model is currently accessible via Tencent Cloud's TokenHub platform, the WorkBuddy and CodeBuddy developer tools, and third-party inference endpoints including OpenRouter and Hermes.

Finance benchmark and hallucination figures

The finance-specific evaluation is the most substantive technical claim in the release. Tencent says it recruited 213 financial professionals, 153 of them from external institutions, to assess 579 real-world tasks spanning fund research, securities analysis and quantitative investment strategy. On an overall performance score, Hy4 returned 3.04, ahead of DeepSeek-V4.1-Flash-1017 (2.97), Kimi K3.1 Pro (2.68) and GLM-5.3 (2.68).

The company also reports that targeted search optimisation reduced Hy4's hallucination rate from 12.95% in its Preview release to 7.2%, while errors from unsupported assumptions fell from 23.1% to 5.6% over the same period. These are internal evaluation figures; the methodology has not been independently audited. Tencent also cites a 12.03 percentage-point improvement over Hy4 Preview in the iRaB benchmark run by financial AI firm Rabyte, which provides one named external reference point.

Beyond finance, the release documents improvements in office productivity (a reported 83.18 average score and 96.15% pass rate on WorkBuddy benchmarks across twelve scenarios), game development (19.1% gain on GameDev, 9.8% on GameLogic versus Preview), and scientific research workflows validated via the TenureTrail benchmark across seven disciplines.

Market context and competitive positioning

The release arrives during a period of intensifying competition among Chinese open-source model developers. Alibaba's Qwen series, Baidu's ERNIE, Zhipu's GLM and the independently backed DeepSeek family are all actively competing for adoption among enterprise developers in Asia and increasingly in Western markets. Tencent's decision to benchmark directly against DeepSeek-V4.1 and GLM-5.3 signals that it regards those as the proximate comparators rather than the US hyperscalers.

The vertical focus on finance is a defensible strategic choice. Regulated financial institutions face strict data-residency and model-governance requirements, and a model pre-tuned on domain-specific tasks may lower integration cost relative to general-purpose alternatives. However, enterprise buyers in this sector will want to see independent red-teaming of the hallucination and assumption-error figures before deploying Hy4 in client-facing or trading-adjacent workflows.

Safety posture and regulatory read-across

Tencent's safety documentation merits attention. The company states that offensive cybersecurity materials are excluded from the training corpus, and that internal assessments by its Xuanwu Lab and Keen Security Lab found no significant concerns about offensive capabilities. On biosafety, the release acknowledges that Hy4's open-ended experimental design capability is more limited than frontier models, which the company frames as a feature rather than a gap.

As an open-weight model, Hy4 falls under the EU AI Act's general-purpose AI rules once it is made available in the European market, including transparency obligations and, depending on capability thresholds, potential systemic-risk designation. The Act's GPAI provisions began applying in August 2025. Tencent has not yet stated whether Hy4 will be submitted for third-party conformity assessment.

Pricing is held flat relative to Hy4 Preview at approximately US$2.70 per million output tokens. A lighter Hy4 Flash variant is scheduled for release imminently, targeting cost-sensitive workloads.

Sector

Topic

Geography