Tencent Hy4: Capabilities, Pricing and Open-Model Plans

Vision and computer use are the practical additions

Tencent says Hy4 can handle fine-grained image processing, complex charts, and mathematical reasoning that uses visual inputs. It also adds computer-use features that let it interact with command-line and graphical interfaces. The company says it trained the model on professional software, including Blender and CAD tools, to support more involved workflows.

If your work depends on information that lives in charts, documents, and interfaces (not just text), this could be genuinely useful. For example: ask the model to interpret a chart, find a related project document, and draft a summary. But a persuasive demo is only half the story. In production, what usually decides success is whether the model can deal with missing or messy inputs, recover from errors, and avoid taking actions you didn't mean to authorize.

Longer tasks need more than a capable model

Long-horizon tasks usually involve multiple decisions: gather information, choose an approach, use tools, inspect results, and revise. Better reasoning helps, but the surrounding system matters just as much. Tool access, state management, task limits, and the ability to pause or undo actions are what determine whether an agent is safe and usable.

Tencent also describes Hy4 as improved at working with long context and following instructions. In a business context, that can help when projects span lots of documents or a long thread of conversation. Still, that doesn't mean teams should assume they can feed "unlimited" context or stop checking outputs. Retrieval quality, source permissions, and information freshness are separate engineering problems that have to be solved regardless.

"Open-source" needs a precise definition

Tencent says it will release Hy4's model weights and system cards in the coming days. Until those materials are available, teams can't independently review the weights, the licensing terms, or the documented limitations. And it's important not to blur "open-source software" with "open-weight models." Releasing weights doesn't automatically mean the training data, code, and full development process are open.

If you're making procurement and technical plans, don't rely on a label - check the actual license and the release documentation. The system card should clarify intended use, known risks, evaluation methods, and safety controls. Those details are what determine whether Hy4 can be self-hosted, modified, and used commercially under the terms your organization needs.

Office work and collaboration

Tencent reports that Hy4 improved the experience in WorkBuddy, its AI-powered workplace product. In WorkBuddy evaluations, it reports an average score of 83.18 and a 96.15% pass rate, with first place in nine of twelve scenarios in the mainland Chinese market. These results are useful for understanding that product's evaluation setup, but they don't automatically generalize across organizations or workflows.

The more practical use case is connecting work context across chats, email, documents, calendars, and meetings through WorkBuddy's integration with WeCom. Tencent says Hy4 can track project history, identify outstanding tasks, flag upcoming meetings that need preparation, and surface deadlines. That's a common pattern in knowledge work: the model helps organize information already spread across tools, instead of forcing staff to copy everything into a prompt.

The trade-off is access. If a system can read messages and documents, it can also see sensitive information. Organizations have to decide which users, projects, and data sources it can access, how permissions are enforced, and whether outputs leak information across teams. Before rolling anything out broadly, the permissions model needs to be tested.

Financial analysis

Tencent says it trained Hy4 on specialized workspaces and tasks based on fund research, securities analysis, and quantitative investment strategies. In a blind evaluation conducted by the Tencent Hy team with Tencent Financial Technology, Hy4 received an overall score of 3.04 across 579 real-world financial tasks. Tencent reports comparison scores of 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for both Kimi K3.1 Pro and GLM-5.3.

Tencent says the evaluation involved 213 financial professionals, including 153 experts from external institutions. It also says Hy4 improved by 12.03 percentage points over Hy4 Preview on an internal Investment Research Agent Benchmark run by financial AI company Rabyte.

This is evidence, not a guarantee that investment decisions will be better. A single aggregate score across tasks can hide where the model is strong or weak, and financial work comes with different risk levels. Drafting a research summary isn't the same as producing a recommendation that moves capital. Before deployment, teams should test Hy4 against their own tasks, require citations or other traceable sources where appropriate, and keep final decisions with accountable professionals.

Software development, games and research

Tencent says Hy4 is built for complex frontend and backend work, including business logic and state management. For game development, Tencent Interactive Entertainment Group developers helped develop the capabilities, and Tencent states support for Unreal Engine, Unity, and Godot. Compared with Hy4 Preview, Tencent reports improvements of 19.1% on its GameDev benchmark and 9.8% on GameLogic.

For developers, the real question is whether generated changes pass the team's tests and fit the architecture. A model that produces a "pretty" interface but mishandles state can create more work than it saves. Use a task set that reflects how your team actually ships: bug fixes, tests, refactoring, documentation, and integration - not just greenfield code generation.

For research, Tencent cites TenureTrail, an evaluation of 123 executable tasks across seven disciplines. It says Hy4 did well in areas including mathematical construction, physical modelling, environmental and earth sciences, systems engineering, and socio-economic decision-making. Tencent also describes work on recursive self-improvement using its Hyra 2.0 harness and says it open-sourced HyCell and HyPDE, models for single-cell representations and physics fields. Those are signals of an ongoing research effort, but teams should still assess each model and tool based on its own documentation and reproducibility.

Treat company-linked evaluations as evidence to validate

Some published results for Hy4 come from evaluations Tencent ran or from evaluations done with Tencent-affiliated partners. The financial blind evaluation included external professionals, which helps, but Tencent still reports the methodology and the aggregate scores. WorkBuddy figures are specific to that product, while the GameDev and GameLogic improvements compare Hy4 to its own preview model.

That doesn't make the results useless. It just means they should motivate a pilot, not replace one. Ask for task-level breakdowns, scoring rubrics, and examples of failure. If a third-party benchmark exists, compare models using the same version, prompts, and tool conditions. A one-point lead often isn't meaningful when task setups differ - or when the higher score required more expensive supervision.

A good pilot evaluation set should include:

• Typical tasks, not just highlight examples.
• Difficult inputs, missing context, and ambiguous instructions.
• A clear definition of what counts as an acceptable answer or action.
• Human review time and correction rates.
• Repeat runs to measure consistency.
• A comparison to the current process, including non-AI options.

Hallucination improvements are not a safety guarantee

Tencent says search optimization reduced Hy4's hallucination rate from 12.95% in Hy4 Preview to 7.2% in internal evaluations. It also says errors from unsupported assumptions dropped from 23.1% to 5.6%. Those numbers suggest improvement in Tencent's testing setup. They don't prove the model will be reliable in every domain, or that it will always cite accurate, up-to-date sources.

Search can improve factual grounding, but it introduces its own failure modes: stale pages, conflicting sources, retrieval gaps, and incorrect synthesis. For high-impact use cases, measure factual errors on your own data. Also make the source trail visible to the reviewer. A lower average error rate only helps if the remaining errors get caught before they cause damage.

Safety controls should match deployment risk

Tencent says Hy4 was trained on defensive cybersecurity data, with proof-of-concept development and vulnerability remediation emphasized, while offensive exploitation materials were excluded. It also says experts from Xuanwu Lab and Keen Security Lab assessed areas like reverse engineering, binary deobfuscation, and vulnerability exploitation, finding no significant concerns in the evaluated areas.

For biological safety, Tencent says internal expert reviews found strong foundational biological knowledge but less capability in open-ended experimental design compared with frontier models. These statements reflect the company's assessment. The full system card and release materials will be important for understanding what the model can and can't handle.

No safeguard at the model level replaces deployment controls. Teams should restrict access to sensitive tools, separate read actions from actions that change systems, maintain audit logs, and set clear escalation paths. The right controls depend on what the model can access. A chatbot that drafts internal notes doesn't need the same controls as an agent that can run commands or modify production environments.

Pricing and the cost of a real task

Tencent lists Hy4 pricing at approximately US$0.90 per million input tokens, US$2.70 per million output tokens, and US$0.045 per million cached input tokens. Tencent says these rates are unchanged from Hy4 Preview despite performance improvements. Hy4 Flash, which is positioned for better cost efficiency, is scheduled for release soon; the announcement doesn't provide its final pricing or capabilities.

Token rates aren't the whole story. A real workflow can involve repeated model calls, retrieval, tool execution, human review, and retries after failures. Longer context can increase input usage, and agentic tasks may require several calls before a job is done. The useful metric is cost per accepted outcome, not cost per request.

A basic pilot should track four things: model and tool spend per task, completion rate without intervention, review time, and the rate of consequential errors. Then compare those results to your current process. If Hy4 cuts drafting time but increases fact-checking, the business case may still work - but the savings may be smaller than the demo implies.

Tencent also announced a temporary promotion: WorkBuddy and CodeBuddy were to be free from November 4 to 15, and Hy4 on TokenHub would be 50% off standard pricing through December 31, 2026. Since promotions and availability can change, confirm current terms before budgeting. You should also check release timing for weights and system cards before planning for open deployment.