Executive summary
In a financial blind test organized by Tencent Hy and Tencent Financial Technology with 213 finance professionals and 579 real tasks, Hy4 ranked first with 3.04, DeepSeek-V4.1-Flash-1017 followed at 2.97, and Kimi K3.1 Pro and GLM-5.3 tied at 2.68. Hy4 leads the third-place models by 0.36 points but leads DeepSeek by only 0.07, so the top two sit in the same tier.
How the comparison was run
The numbers come from the financial blind evaluation Tencent published on November 4, 2026, alongside the Hy4 release. Tencent Hy and Tencent Financial Technology organized the test. The panel comprised 213 finance professionals, 153 from leading external institutions, with about 70% holding more than five years of industry experience. The task set covered 579 real financial scenarios.
The scope matters. This is a vendor-commissioned and vendor-organized blind test. The press release does not publish the full task list, the scoring rubric or confidence intervals. The table below therefore describes the relative order under Tencent's methodology; it is not a measure of absolute capability across all financial work.
Tencent also reported two longitudinal improvements for Hy4 against its own predecessor, Hy4 Preview: a 12.03 percentage-point gain on the internal Investment Research Agent Benchmark (iRaB), and a fall in hallucination rate from 12.95% to 7.2%. Errors caused by unsupported assumptions dropped from 23.1% to 5.6%. These compare Hy4 with Hy4 Preview, not with the other three models.
The four models
Tencent Hy4. Released on November 4, 2026, with 770B parameters. It adds visual understanding and computer use, is available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy and Yuanbao, and is scheduled to receive an open-weight release with a system card.
DeepSeek-V4.1-Flash-1017. Ranked second in the blind test with 2.97. The release does not disclose its parameter count, context length or training-data scope, and does not state whether the evaluated version uses open weights.
Kimi K3.1 Pro. A model in Moonshot AI's Kimi series, scored 2.68. Parameter count and version details were not disclosed.
GLM-5.3. A model in Zhipu's GLM series, also scored 2.68 and tied with Kimi K3.1 Pro. Parameter count and version details were not disclosed.
The financial blind-test score is the only metric measured on a common basis for all four models. Tencent disclosed parameter count only for Hy4. Inference cost, context length and tool-calling performance were not reported on a comparable basis, so this comparison does not rank those dimensions.
Core comparison table
| Model | Financial blind-test score | Difference vs Hy4 | iRaB vs predecessor | Parameters | Open weights |
|---|---|---|---|---|---|
| Tencent Hy4 | 3.04 | Baseline | +12.03 pp | 770B | Release planned |
| DeepSeek-V4.1-Flash-1017 | 2.97 | -0.07 | Not disclosed | Not disclosed | Not stated |
| Kimi K3.1 Pro | 2.68 | -0.36 | Not disclosed | Not disclosed | Not stated |
| GLM-5.3 | 2.68 | -0.36 | Not disclosed | Not disclosed | Not stated |
Test design
| Element | Detail |
|---|---|
| Organizers | Tencent Hy team, Tencent Financial Technology |
| Method | Blind evaluation |
| Participants | 213 finance professionals |
| External experts | 153, from leading external institutions |
| Experience | About 70% with more than five years in the industry |
| Task count | 579 real financial tasks |
| Task types | Fund research, securities analysis, quantitative investment strategies |
| Disclosure | Composite scores only; no sub-scores or error margins |
Key differences
The tier boundary is 0.36 points. Hy4's 3.04 and DeepSeek-V4.1-Flash-1017's 2.97 are separated by 0.07 points. Both are 0.36 points ahead of Kimi K3.1 Pro and GLM-5.3 at 2.68. Under Tencent's ordering, Hy4 and DeepSeek form the first tier, while Kimi and GLM tie behind them.
Hy4's generational gain is larger than its cross-model lead. Against Hy4 Preview, Hy4 gained 12.03 points on iRaB, cut hallucination by 5.75 points and cut unsupported-assumption errors by 17.5 points. Against DeepSeek in the blind test, its lead is 0.07 points. The two findings describe different things: strong improvement over its own prior version, limited separation from its closest competitor.
Financial tasks test a different mix than general benchmarks. Tencent says the tasks centered on fund research, securities analysis and quantitative strategies, with finance professionals as evaluators. That is closer to real workflows than a general reasoning benchmark, but participant composition, task difficulty and scoring consistency all affect the result. The release does not report inter-rater agreement.
Kimi K3.1 Pro and GLM-5.3 tie without sub-scores. Both scored 2.68. Without category-level results, their relative strength in report generation, data verification or strategy design cannot be determined.
Deployment conditions are not on the same footing. Hy4 is slated for an open-weight release at 770B parameters. The other three models' weight availability and parameter counts were not stated in this announcement. For institutions that need private deployment, weight availability, GPU memory requirements and inference-framework support often matter more than a 0.07-point gap.
Conclusion and selection guidance
Under the financial blind test Tencent published, Hy4 ranks first, DeepSeek-V4.1-Flash-1017 is a close second, and Kimi K3.1 Pro and GLM-5.3 tie for third. That ordering applies to this task set and rubric; it should not be generalized into a universal ranking across all financial use cases.
For finance-heavy work inside the Tencent ecosystem, Hy4 is the first option under this methodology. It is already available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy and Yuanbao, and supports visual understanding and computer use, which suits tasks that require reading charts and executing long multi-application workflows. At 770B parameters, local deployment is expensive, so API access is the near-term route.
For cost-sensitive or open-source-first teams, DeepSeek-V4.1-Flash-1017 is a same-tier alternative. A 0.07-point difference is not enough on its own to decide a deployment. Compare cost per token, latency, context length, tool-calling stability and compliance requirements instead.
For long-document research or Chinese knowledge work, Kimi K3.1 Pro and GLM-5.3 remain viable. Both scored 2.68, 0.36 points behind the leaders. Suitability depends on the specific sub-task and on private-deployment conditions, not on the composite score alone.
Treat the result as vendor-reported until independently replicated. Tencent Hy and Tencent Financial Technology organized the test, and the full task list, rubric and error margins are not public. Before procurement or rollout, institutions should run a small paired test on their own business data and record the task set, scoring rules and model versions.
Data sources and references
- Tencent official press release (PR Newswire): Tencent Launches Hy4: Joining the Ranks of the World's Leading Open-Source AI Models
- AAP reprint: Tencent Launches Hy4
- TechNode: Tencent launches Hy4 with vision and computer-use capabilities