A model that can generate code is useful. A model that can inspect a computer screen, interact with applications and execute a sequence of actions has a much broader set of possibilities.

It also creates a different security problem.

Tencent's new Hy4 release illustrates both sides of that development. The company is expanding the capabilities of its flagship AI system while introducing restrictions that it says are intended to reduce the risk of misuse.

The announcement comes as advanced AI models are increasingly incorporated into development tools, office software and autonomous workflows.

Those environments can give models access to files, applications and external services. Their potential impact therefore depends on more than the quality of the text they generate.

Tencent is positioning Hy4 as a general-purpose model designed for practical work, with additional safeguards covering cybersecurity and biological risks.

From long-context text processing to visual understanding

Tencent released Hy4 Preview on August 28, 2026.

The preliminary model used a mixture-of-experts architecture with 770 billion total parameters and approximately 49 billion activated during each token-processing step.

This design allows the system to use selected parts of its overall parameter capacity rather than performing every calculation through the complete network.

The preview also offered a context window exceeding one million tokens, supporting tasks involving substantial documents, codebases and extended interactions.

The November version introduces visual understanding, allowing Hy4 to analyze images and extract information from visual material.

Tencent identifies complex chart interpretation, image analysis and mathematical reasoning involving visual information as important applications.

The company also reports improvements in instruction following, long-context comprehension and tasks requiring extended sequences of decisions.

These capabilities do not guarantee flawless execution.

A model may correctly identify elements in a document while still making mistakes when combining information or deciding which action to perform next.

The reliability of visual reasoning must therefore be distinguished from the availability of the feature itself.

Computer use expands what the model can actually do

Hy4 now supports computer interaction through both graphical user interfaces and command-line tools.

This is a substantial step beyond generating instructions for a human operator.

An AI agent with appropriate tools can inspect an interface, decide which operation to perform and continue working after receiving the result.

Tencent says it has provided specialized training for professional software environments, including Blender and computer-aided design applications.

The company highlights long-horizon tasks that require several coordinated operations rather than a single response.

These capabilities could support software development, document processing and other workplace activities.

They also increase the importance of access controls.

An agent that misunderstands an instruction but has permission to modify important files may cause damage even without malicious intent.

The same concern applies to actions involving external systems, account permissions or confidential information.

Consequently, an AI model's safety training cannot replace application-level authorization and operational supervision.

Tencent says it filtered offensive cybersecurity material

A central part of the announcement concerns how Hy4 was trained.

Tencent says it prioritized cybersecurity material associated with defensive activities, including understanding and remediating software vulnerabilities.

The intention is to preserve useful technical knowledge while reducing the model's ability to assist in offensive attacks.

However, the exact scope of the restriction is not entirely clear from the official descriptions.

Some materials associated with the Hunyuan team describe reverse engineering and penetration testing as categories excluded from training alongside offensive exploit development.

Tencent's main release offers a different account.

It states that reverse engineering and penetration testing were incorporated for analytical purposes, while offensive exploitation material was excluded.

Those statements cannot simply be treated as identical.

Reverse engineering can help security researchers understand malicious software, identify vulnerabilities or investigate how a program operates.

Penetration testing can be a legitimate, authorized process for identifying weaknesses in systems.

Both areas also involve techniques that could potentially be used for malicious purposes.

The difference between defensive research and offensive activity often depends on the intended use, operational context and authorization.

The available announcements do not fully explain the criteria used to classify the training material.

It would therefore be inaccurate to claim that Hy4 received no exposure to reverse-engineering or penetration-testing content.

The supported conclusion is narrower: Tencent says it attempted to reduce offensive cybersecurity capabilities by restricting certain training materials while retaining knowledge useful for defensive analysis.

Internal security assessments are not independent guarantees

Tencent says two security teams, Xuanwu Lab and Keen Security Lab, evaluated Hy4 before release.

The assessments examined areas including cryptographic analysis, reverse engineering, binary deobfuscation and vulnerability exploitation.

According to Tencent, the results did not reveal significant offensive cybersecurity concerns within the tested categories.

That is a company-reported assessment.

The teams operate within Tencent's research and security ecosystem, and the detailed evaluation conditions have not yet been fully disclosed through the promised system documentation.

A model performing relatively poorly on a particular offensive benchmark does not establish that it cannot assist with harmful activities in other circumstances.

Results depend on the test design, the prompts used and the criteria for determining whether a dangerous capability was demonstrated.

The risks can also change when a model receives access to external tools.

Computer-use permissions, command execution and access to networked resources can increase the consequences of an incorrect or unauthorized decision.

An effective deployment therefore requires safeguards beyond the model's training data.

These can include restricted permissions, action logging, approval requirements and mechanisms for interrupting automated workflows.

Tencent's announced restrictions should be understood as one layer of risk reduction rather than a comprehensive guarantee.

Biological safety is part of the same policy

Tencent also says it evaluated Hy4's potential to support harmful biological research.

The company's internal experts found that the model retained substantial general biological knowledge but showed more limited abilities in open-ended experimental design.

Tencent compares those capabilities with more advanced models in the field.

This should not be interpreted as proof that Hy4 cannot produce any sensitive biological information.

The findings describe performance under selected evaluation conditions.

The system card promised alongside the model weights may provide further information about the tasks, criteria and thresholds used.

Independent assessments will remain necessary to establish whether the reported safeguards continue to perform under different conditions.

The service is live, but the final weights are not yet public

Tencent's announcement repeatedly refers to open-source AI.

An important distinction is required when interpreting that description.

Hy4 is already accessible through several products and services.

Tencent identifies WorkBuddy, CodeBuddy, Yuanbao, ima, WeCom's Dayuan assistant and Tencent Cloud TokenHub among the available entry points.

API access is also being offered through third-party platforms including OpenRouter, Hermes and OpenCode.

These services allow developers and users to interact with a hosted model.

They do not automatically provide access to the model's weights or the complete details required to run it independently.

Tencent says it intends to publish the full Hy4 model weights and its system card in the days following the November 4 announcement.

Hy4 Preview had previously been released with open weights in August.

That earlier release should not be confused with the availability of the final November version.

At the time of the new announcement, the company was promising further publication rather than confirming that all updated files were already accessible.

The distinction matters for researchers interested in inspecting the model, reproducing evaluations or running it outside Tencent's hosted infrastructure.

Full access will also require examining the license and accompanying technical documentation.

Pricing places Hy4 in the competitive API market

Tencent has disclosed commercial API pricing for the new model.

The company quotes approximately $0.90 per million input tokens and $2.70 per million output tokens.

Cached input is listed at approximately $0.045 per million tokens.

Tencent says this preserves the price level of Hy4 Preview despite the additional capabilities.

Actual charges may differ across providers, currencies and promotional programs.

For Tencent Cloud TokenHub, the company is offering a 50% promotional discount through December 31, 2026.

WorkBuddy and CodeBuddy are also included in a promotional free-access period scheduled for November 4 through November 15, subject to the relevant service conditions.

A more cost-focused version called Hy4 Flash is planned for a future release.

That variant should not be considered generally available merely because it was mentioned alongside the main model.

Tencent's benchmarks show progress, but their scope is limited

The company has published several sets of performance results.

In WorkBuddy-related office and collaboration tasks, it reports an average score of 83.18 and a 96.15% pass rate.

Tencent says Hy4 led nine of twelve evaluated scenarios in the mainland Chinese market.

A separate financial assessment involved 213 professionals reviewing 579 real-world tasks.

According to Tencent, Hy4 achieved a score of 3.04 and ranked ahead of the selected competing models.

These findings do not represent a universal ranking of all available AI systems.

The result depends on which tasks were chosen, which models participated and how the evaluation was conducted.

Tencent also reports improvements in game development.

Its GameDev and GameLogic benchmarks show gains of 19.1% and 9.8%, respectively, compared with Hy4 Preview.

The company has described work involving Unreal Engine, Unity and Godot.

Those results do not establish an equivalent improvement in the speed or quality of every real-world game development project.

Reliability results require additional care.

Tencent's main announcement reports a decline in hallucination rates from 12.95% to 7.2% in selected internal search evaluations.

Other Tencent ecosystem materials refer to a decline from 15.76% to 5.45%.

The differences cannot be reconciled into one universal hallucination rate without knowing the precise test conditions.

The figures should remain attached to their reported evaluation contexts rather than being presented as an overall probability that Hy4 produces inaccurate information.

Tencent is building AI around its existing products

Hy4 is being developed alongside applications rather than as a purely standalone model.

WorkBuddy and CodeBuddy are particularly important examples of that approach.

Tencent uses these products to develop model capabilities around office workflows, software engineering and interactions with external tools.

WorkBuddy can connect with workplace information, documents, messages and calendars through supported services and configurations.

The company describes AI assistance for tracking projects, preparing meetings and prioritizing work.

Other Tencent services are expected to adopt Hy4 capabilities over time.

QQ, Honor of Kings, Game of Peace, WeRead and Tencent News are among the products named in the company's integration plans.

These references do not mean that every Hy4 feature is already available in each application.

Deployment conditions and supported functions can differ significantly between products.

The strategic objective is to connect model development with Tencent's existing cloud, communications, gaming and business software ecosystem.

Security claims face a more demanding test as agents gain autonomy

The most consequential aspect of Hy4 may be the combination of visual understanding, extended planning and direct computer interaction.

A model that produces an inaccurate answer may mislead a user.

An agent with permission to operate software can potentially turn an inaccurate decision into a change to files, applications or external systems.

This makes security controls particularly important.

Tencent's attempt to reduce offensive capabilities through training-data selection addresses one component of the problem.

However, it does not eliminate risks arising from permissions, tool integrations, incorrect reasoning or unexpected instructions encountered during automated workflows.

Organizations deploying such systems still need operational safeguards and independent validation.

Tencent says internal security assessments form a required step before releasing advanced models openly.

The promised publication of the full Hy4 weights and system card should provide additional material for examining that process.

The discrepancies between official descriptions of included and excluded cybersecurity data also make detailed disclosure especially important.

For now, the confirmed position is that Hy4 launched on November 4 through several hosted products and APIs, with new visual and computer-use capabilities.

Its expanded open-weight release and full safety documentation were still scheduled for subsequent publication at the time of the announcement.

The effectiveness of Tencent's restrictions against offensive misuse remains a company claim awaiting broader independent examination.