Tencent Hy4 debuts with native vision and computer use, tops open-weight financial benchmarks

Hy4 is Tencent's open-weight frontier model that, for the first time in the Hy series, combines native visual understanding with computer-use agentic capabilities, with downloadable weights and system cards promised in the days after its Nov. 4, 2026 release.

Tencent released Hy4 on Nov. 4, 2026, moving the model from the August preview into general availability with first-time visual understanding and computer-use capabilities. The company said the release places Hy among the strongest open-source models, and it promised downloadable weights and system cards within days. Hy4 is available immediately through Tencent Cloud TokenHub and OpenRouter, and it is integrated into WorkBuddy, CodeBuddy, Yuanbao and ima. Agentic Weekly's November model tracker corroborates the launch date and the model's 1 million-token context.

The release extends the model line Tencent has built under the Hunyuan brand, where earlier iterations concentrated on text generation and reasoning. Hy4 adds native vision and agentic computer use in a single open-weight system, a combination that has been rare outside closed frontier models. Tencent said the new model supports fine-grained image processing, complex chart analysis and mathematical visual reasoning, and that its computer-use training covers command-line interfaces, graphical user interfaces and specialized software such as Blender and CAD.

What is new in Tencent Hy4?

The most consequential change is visual understanding. Hy4 can process images as part of its core architecture rather than through a separate vision module, which Tencent said improves performance on tasks that require reading dense charts, interpreting screenshots and performing math on visual inputs. That capability is paired with a 1 million-token context window, allowing the model to hold long documents, codebases and multi-step tasks in a single session.

The second major addition is computer use. Hy4 is trained to operate command-line interfaces and graphical user interfaces, meaning it can execute multi-step workflows that move between terminals, desktop applications and web tools. Tencent said the training included specialized software such as Blender and CAD, an indication that the agentic layer is aimed at professional production environments rather than simple web automation.

Tencent said Hy4 supports vision perception for the first time, with fine-grained image processing, complex chart analysis and mathematical visual reasoning. The capability is paired with a 1 million-token context window, allowing the model to reason about an image and a long text document in the same session, a requirement for agentic tasks that involve reading a screen and then acting on it.

Computer-use models are evaluated not only on whether they produce a correct action but on whether they can recover from errors, re-read a screen and maintain a plan across many steps. Tencent's training on Blender and CAD is a signal that the company measured success against professional software, where a single misclick can corrupt a file. The model's 1 million-token context gives it room to retain a full session transcript, including tool outputs and error messages.

1. Native visual understanding for fine-grained image processing, complex chart analysis and mathematical visual reasoning.
2. Computer-use training for command-line interfaces, graphical user interfaces and specialized tools including Blender and CAD.
3. A 1 million-token context window for long-context comprehension and long-horizon agentic tasks.
4. Availability through Tencent Cloud TokenHub, OpenRouter, WorkBuddy, CodeBuddy, Yuanbao and ima.
5. A commitment to release open weights and system cards in the coming days.

How does Hy4 perform on financial and agent benchmarks?

Tencent evaluated Hy4 on 579 real-world financial tasks with 213 finance professionals, including 153 experts from leading institutions and more than 70% with over five years of experience. Hy4 scored 3.04 overall, the highest among open-weight models in the evaluation. DeepSeek-V4.1-Flash-1017 scored 2.97, and Kimi K3.1 Pro and GLM-5.3 each scored 2.68.

Tencent's 579-task financial evaluation of open-weight models

Model	Overall score
Hy4	3.04
DeepSeek-V4.1-Flash-1017	2.97
Kimi K3.1 Pro	2.68
GLM-5.3	2.68

Tencent also reported a 12.03 percentage-point improvement over Hy4 Preview on the internal Investment Research Agent Benchmark, or iRaB. The gain reflects the model's ability to handle long-horizon research tasks, where the combination of vision, longer context and agentic tool use reduces the number of handoffs a financial workflow requires.

Today, we're releasing Hy4, another major leap forward that places Hy among the strongest open-source models for the first time.Tencent Hy Team

An independent evaluation from RabyteTech, a financial AI agent company, placed Hy4 in a tie with Claude Opus 5.5 for second place overall among all models tested. RabyteTech said Hy4 scored 75.4 on data queries, nearly 20 points ahead of GPT-6.1 Sol and first among all evaluated models. The tie is notable because Hy4 is an open-weight model competing against closed frontier systems.

Hy4 is currently the only Chinese model to tie with Claude Opus 5.5 for second place overall. On data queries, it scored 75.4—nearly 20 points ahead of GPT-6.1 Sol—ranking first among all models evaluated.RabyteTech

On WorkBuddy, Tencent's agent benchmark for business software tasks, Hy4 recorded an average performance score of 83.18 and a pass rate of 96.15%. It ranked first in nine of twelve scenarios in the mainland Chinese market. The benchmark measures how reliably a model completes tasks inside business applications, and the results suggest Hy4 is closer to production-ready agent behavior than its preview predecessor.

WorkBuddy is Tencent's agentic office product, and its benchmark scenarios span document editing, spreadsheet manipulation, email drafting and presentation building. The 96.15% pass rate means Hy4 completed tasks without human intervention in the large majority of cases, while the 83.18 average score captures the quality of the output, not just whether a task finished. First place in nine of twelve mainland China scenarios is a measure of reliability across different software environments.

Where can developers run Hy4 and what does it cost?

Hy4 is available through the Tencent Cloud TokenHub API and on OpenRouter, alongside other aggregators, with a 1 million-token context window. It is also embedded in Tencent products: WorkBuddy for agentic office work, CodeBuddy for software development, Yuanbao for general assistance and ima for knowledge management. Tencent said API pricing is unchanged from the August preview.

The listed price is approximately $0.834 to $0.90 per million input tokens, $2.501 to $2.70 per million output tokens, with a lower rate for cached input. That pricing places Hy4 in the same range as other frontier open-weight models and makes vision and computer-use capabilities available without a premium over the text-only preview.

At the listed rates, Hy4 is priced competitively with DeepSeek and GLM models, which have made open-weight pricing a central part of their distribution strategy. Tencent's decision to hold the line at preview pricing while adding vision and computer use is a bid to keep the total cost of an agentic workflow predictable. Cached input pricing, which is lower than standard input pricing, is designed for workloads that repeatedly process the same documents.

Why do open weights matter for Hy4 adoption?

Tencent said it plans to release model weights and system cards in the coming days, which would make Hy4 the first open-weight model to combine native vision with full computer-use agentic capabilities. The company framed the release around responsible open-source practices, with explicit emphasis on cybersecurity and biological safety in the system card review process.

Open weights change how enterprises can deploy the model. Teams can fine-tune Hy4 on proprietary data, run it in air-gapped environments and inspect its behavior directly rather than relying on API-level observability. For financial institutions with data-residency requirements, that distinction can be decisive.

The move also intensifies competition among Chinese open-weight labs. DeepSeek and Zhipu AI already ship strong open models, and Tencent's financial-task lead gives it a differentiated position. At the same time, Hy4's tie with Claude Opus 5.5 in RabyteTech's evaluation puts pressure on closed frontier vendors to justify API-only distribution.

The promise of open weights also carries governance questions. Computer-use models can be fine-tuned to automate tasks that range from benign office work to malicious activity, and Tencent's emphasis on cybersecurity and biological safety acknowledges that risk. The system card, when released, will document the model's behavior on safety evaluations and should give enterprises a basis for risk assessment.

Tencent's position is unusual among open-weight labs because the company controls both the model and the major distribution products. A developer can test Hy4 on OpenRouter today, move to Tencent Cloud for production, and eventually run the same weights in a private environment. That continuity lowers switching costs and gives Tencent a commercial path even if a competitor matches the model's benchmark scores.

What do the benchmark results mean for financial and enterprise AI?

Financial services is the clearest early use case. Hy4's 3.04 score on 579 real-world tasks covers data queries, document analysis, chart reading and investment research, and the 1 million-token context allows an analyst to load an entire research packet into a single session. The 12.03-point iRaB gain over Hy4 Preview indicates that long-horizon research workflows are where the new capabilities compound.

Enterprise deployments are likely to start with WorkBuddy and CodeBuddy, where Hy4's computer-use training can be applied to office automation and development tooling. The specialized training on Blender and CAD suggests Tencent is also targeting engineering and design workflows, where GUI agents have historically struggled. Tencent Financial Technology, the company's payments and financial services unit, could be among the first internal divisions to apply Hy4's financial-task gains, though Tencent has not announced a specific rollout.

Financial institutions are often restricted from sending data to third-party APIs, which has slowed adoption of closed frontier models in banking and insurance. An open-weight model with Hy4's financial-task scores removes that barrier. The 1 million-token context is particularly valuable in financial research, where a single research report can be long enough to strain shorter context models, and where regulators expect institutions to retain records of model inputs and outputs.

What comes next for Hy4 and the Hy series?

The immediate next step is the release of weights and system cards. Tencent said the documentation would emphasize responsible open-source practices, focusing on cybersecurity and biological safety, an indication that the company is seeking to preempt safety objections as it distributes a model with computer-use capabilities.

After that, the trajectory depends on how the open-weight community adapts Hy4. If developers fine-tune the model for vertical domains, the vision and computer-use features could become the base layer for a new generation of agentic products. Tencent's product integrations give it a distribution channel that most open-weight labs do not control, and its financial benchmark leadership gives it a credible enterprise narrative.

The open-weight frontier has narrowed in 2026, with most labs converging on long context and agentic tool use. Hy4's combination of native vision and computer use is the differentiator Tencent can claim until a competitor ships a comparable open model. The company's cloud and product distribution gives it an enterprise path that pure model labs lack, and the financial benchmark leadership gives procurement teams a concrete evaluation point.

Tencent has not said whether a larger Hy variant is planned, but the company's release cadence since August suggests updates will continue. The focus on financial benchmarks and professional software indicates that Tencent is prioritizing enterprise revenue over consumer chatbot adoption, a strategy that aligns with the capabilities Hy4 introduced.

What questions do developers ask about Hy4?

When will Hy4 weights be released? Tencent said it plans to release model weights and system cards in the coming days, with responsible open-source practices focused on cybersecurity and biological safety.

Where can developers access Hy4? The model is available via API through Tencent Cloud TokenHub and OpenRouter, and inside WorkBuddy, CodeBuddy, Yuanbao and ima.

How much does Hy4 cost? Pricing is unchanged from the preview: approximately $0.834 to $0.90 per million input tokens and $2.501 to $2.70 per million output tokens, with a lower rate for cached input.