Tencent Launches Hy4 With Vision and Computer Use, but Weights and System Cards Come "in the Coming Days"; Benchmark Figures Come From the Release Itself, and Prices Differ From the August Preview Post

On November 4, 2026, Tencent published a press release on its website announcing Hy4, its latest-generation large language model. The release says the model ranks among the world's leading open-source AI models. In the same release, Tencent Hy says it plans to publish model weights and system cards "in the coming days". So on the day of the announcement the weights and system cards were not out, and "open-source" describes a plan, not yet a fact. The release gives no parameter count or context length. For those, only the August 28 preview post offers numbers: 770B total parameters, 49B active parameters and a context window above 1M tokens, and preview figures should not be read as figures for the final model.

On capability, the release says Hy4 adds visual understanding for the first time, covering fine-grained image processing, complex chart analysis and mathematical visual reasoning. It also adds computer use for the first time, working with command-line and graphical interfaces, with specialized training on professional software such as Blender and CAD. For office work, an evaluation on Tencent's own WorkBuddy product, with no evaluator named, reports an average score of 83.18 and a pass rate of 96.15%, with first place in nine of twelve scenarios in the mainland Chinese market. For finance, the Tencent Hy team and Tencent Financial Technology ran a blind evaluation with 213 financial professionals, 153 of them external experts and about 70% with more than five years of experience, on 579 real tasks. Scores were Hy4 3.04, DeepSeek-V4.1-Flash-1017 2.97, Kimi K3.1 Pro 2.68 and GLM-5.3 2.68. The release does not state the maximum score, so the ranking is clearer than the absolute level. The release adds that Hy4 connects to WeCom through WorkBuddy and draws on chats, emails, documents, calendars and meetings, tracking project history across several projects and flagging open tasks, meetings that need preparation and approaching deadlines. New team collaboration features powered by Hy4 are to arrive with WorkBuddy's upcoming "Projects" feature.

The preview post offers a second yardstick, and it is not the same test. For Hy4 preview, Tencent ran an internal blind evaluation with 163 experts on 203 engineering tasks and reported an average of 2.99 out of 4.00, slightly ahead of GLM-5.3 at 2.92 and Kimi K3 at 2.94. The final release's finance score of 3.04 comes from 579 finance tasks, and its comparison set uses other model versions such as DeepSeek-V4.1-Flash-1017 and Kimi K3.1 Pro. The tasks, raters and rivals differ, so the step from 2.99 to 3.04 is not a measured improvement. The preview post also says the preview model helped optimize its own inference system and raised end-to-end throughput by 31.8% over a baseline, while the final release says Hy4 was used to optimize pre-training infrastructure for smaller models. Both are Tencent's own account.

The release also gives internal comparisons against Hy4 Preview. On the iRaB investment research agent benchmark from the financial AI company Rabyte, Hy4 improved by 12.03 percentage points. After search optimization, the hallucination rate fell from 12.95% to 7.2%, and errors caused by unsupported assumptions fell from 23.1% to 5.6%. On the GameDev benchmark Hy4 improved by 19.1% and on GameLogic by 9.8%. These are results against the preview in internal tests. The release gives no test-set sizes or methods for them, and no like-for-like comparison with third-party models. For science, it cites TenureTrail, 123 executable tasks across seven disciplines. It also says Hy4 used the Hyra 2.0 harness to optimize pre-training infrastructure for smaller models with the same architecture, and that the team open-sourced two foundation models, HyCell for single-cell representations and HyPDE for physics fields.

The safety section deserves its own reading. On cybersecurity, Tencent says Hy4 was trained on defensive security data that favors proof-of-concept development and vulnerability remediation, with reverse engineering and penetration testing included for analysis, while offensive exploitation material was explicitly excluded. Experts from Tencent's Xuanwu Lab and Keen Security Lab ran assessments and found no significant concerns about offensive capability in the areas evaluated. On biosafety, internal expert reviews found strong foundational biology knowledge but more limited open-ended experimental design than frontier models. These are assessments of the areas evaluated and internal reviews, not independent audits, and the system card is not yet out.

On availability, Hy4 is offered through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, the WeCom AI assistant Dayuan, ima and more, and its API is being made available on third-party platforms such as OpenRouter, Hermes and OpenCode. The release says pricing is the same as for the preview, at about US$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens. The August 28 preview post, however, states USD 0.834, 2.501 and 0.042. The two official texts do not match, and they are shown here side by side without an explanation that Tencent has not given. A section heading also says Hy4 Flash was "recently released", while the text says it is "scheduled for release soon". WorkBuddy and CodeBuddy are free from November 4 to 15, and Hy4 on TokenHub is half price through December 31, 2026.

For readers this is a vendor launch release. The weights are not out, all evaluation figures come from the release itself; the finance blind test was run by Tencent teams, while the WorkBuddy evaluator and its independence are not disclosed, comparison sets and score scales are incomplete, and the prices do not agree between two Tencent pages. To judge whether Hy4 really belongs among the leading open-source models, wait for the weights, the system card and independent third-party tests.