Tencent says the stable release of Hy4 improved on Hy4 preview across multiple internal evaluations and product scenarios, and is now available through channels including Tencent Cloud and OpenRouter. The page lists several quantitative results but does not fully disclose test sets, baseline configurations, or evaluation procedures, so the results should not be taken to predict performance on every task.
Suitable for assessing: The capability areas Tencent reports for the stable release of Hy4, specific results reported by Tencent and its partners, launch channels, and API pricing.
Not suitable for extrapolating: Performance on undisclosed benchmarks, success rates across all user tasks, or direct comparisons with models evaluated using different datasets and procedures.
Model version covered: Stable release of Hy4; most comparison baselines in the article are Hy4 preview.
Test environment or client: Tencent internal tests, Tencent product scenarios, and partner-run evaluations; specific model parameters and runtime environments are often not stated.
Inference tier and parameters: Not specified.
The launch page summarizes evaluations by Tencent's internal and product teams and also cites results from partners in finance and investment research. The quantitative evaluations visible on the page include a finance crowd evaluation, the internal IRAB, the TenureTrail research tasks, WorkBuddy's overall evaluation, and product scenarios in Yuanbao, Honor of Kings, QQ, and other products. Disclosure of each evaluation's task composition, sample selection, scoring criteria, number of runs, and statistical uncertainty varies. The results below are recorded as reported in the original and are not combined into a single benchmark.
Finance crowd evaluation: Tencent says 213 experts participated, including 153 from leading external financial institutions, and about 70% had more than five years of work experience. On 579 real-world financial tasks, Hy4 scored 3.04/4, ahead of DeepSeek-V4.1-Flash-1017 (2.97/4), Kimi K3.1 Pro (2.68/4), and GLM-5.3 (2.68/4), as listed on the page. The original says Hy4 ranked first among open-source models overall.
Financial search and factual consistency: In Tencent's in-house test, Hy4's hallucination rate fell from 15.76% to 5.45% compared with Hy4 preview, while the error rate attributed to guessing fell from 23.1% to 5.6%. The page does not provide the test set size or metric definitions.
Xuntu Technology IRAB: The partner says it evaluated the stable release of Hy4 on 101 real-world investment research tasks. Hy4 received an overall score of 61.63, up 12.03 points from Hy4 preview. The page does not state the maximum score, task composition, or scoring criteria.
Research tasks: Tencent says its internal TenureTrail evaluation covered 123 executable research tasks across 7 disciplines. Hy4 ranked first among open-source models in mathematical construction, physical modeling, Earth and environmental science, systems engineering, and socioeconomic decision-making. The original does not provide scores by discipline.
WorkBuddy: The page reports an average overall evaluation score of 83.18 and a pass rate of 96.15%. Hy4 ranked first among domestic models in 9 of 12 scenarios. The page provides no details about the samples or scoring.
Yuanbao: The internal evaluation score rose from 65.25 for Hy4 preview to 72.53 for Hy4. Multi-turn memory improved by 6.67 points, and writing diversity by 10.89 points. Task completion for multimodal agents improved 36.8%, and multimodal question answering improved 15.8%; the original does not further explain how these percentages were calculated.
Honor of Kings: Accuracy on 20 types of game configuration scenarios rose from 91.74% to 97.28%, and reached 99.32% after autonomous verification and repair by an agent. Accuracy in root-cause analysis of bugs rose from 59.5% to 73.8%. The original says tool-call depth was 3.7 times that of preview, with no notable increase in latency or token consumption.
Other product scenarios: Human ratings for the AI NPC in Game for Peace averaged 2.91/3, with an emotional value subscore of 2.96/3. Overall satisfaction on QQ rose from 90.11% to 93.10%. Instruction following in WeCom improved 13.5%, complex information retrieval improved 12.7%, and overall message and conversation performance improved 8.5%. A blind test by Tencent News reported a 36% improvement in end-to-end answer quality. These results use each product's own evaluation criteria and are not comparable across products.
Safety evaluation: Hy-BioExperimentBench gave Hy4 a score of 30.4/100 for open-ended experimental design. The page says Tencent's internal safety audit found no notable risks in Hy4's offensive capabilities related to cryptographic algorithm reverse engineering, adversarial binary obfuscation, or vulnerability exploitation. These are conclusions from evaluations disclosed by Tencent.
| Item | Value disclosed on the page | Scope or comparison stated on the page |
|---|---|---|
| Finance crowd evaluation | Hy4 3.04/4; DeepSeek-V4.1-Flash-1017 2.97/4; Kimi K3.1 Pro 2.68/4; GLM-5.3 2.68/4 | 579 tasks; 213 experts participated |
| Hy4 hallucination rate | 15.76% → 5.45% | Hy4 preview → Hy4; Tencent in-house test |
| Error rate attributed to guessing | 23.1% → 5.6% | Hy4 preview → Hy4; Tencent in-house test |
| TenureTrail | 7 disciplines, 123 tasks | Tencent internal evaluation; scores not disclosed |
| WorkBuddy | Average performance 83.18; pass rate 96.15% | 12 scenarios; ranked first among domestic models in 9 |
| Yuanbao internal set | 65.25 → 72.53 | Hy4 preview → Hy4 |
| Honor of Kings configuration accuracy | 91.74% → 97.28% → 99.32% | 20 types of scenarios; the last figure is after autonomous verification and repair by an agent |
| Hy4 API pricing | Cached input ¥0.3; input ¥6; output ¥18 | Per million Tokens |
The page says Hy4 is available on Tencent Cloud and OpenRouter, and can also be tried in products such as WorkBuddy/CodeBuddy, Yuanbao, and ima. The links on the page show Tencent Cloud's modelId=hy4 and the OpenRouter page path tencent/hy4. See the original for API pricing and stable-release information.
This is vendor-published material and is useful for confirming the official release date, availability, public pricing, and evaluation results disclosed by Tencent. Some partner results come from their own businesses or evaluation frameworks, and many metrics lack complete methodological details. Claims such as “ranked first” and “improved” apply only to the samples, scenarios, and criteria specified in the original. The launch page also acknowledges that Hy4 still trails leading closed-source models, particularly on difficult internal tests involving frontier mathematics, long-horizon development, model iteration, and visual question answering. Scores and percentages from different evaluations are not directly comparable and do not guarantee performance in arbitrary use cases.
Open the Tencent Hunyuan official source in Tabbit and verify the title, publication date, body text, and channel links listed directly on the page. This record was collected on 2026-11-16. The launch page does not provide the complete test sets, prompts, model parameters, or scoring code needed to reproduce most of the results above. This article therefore records the reported results as presented and does not claim independent reproduction.
Hy4