Tencent Hy4 Launches with Vision and Computer Use: What Developers Need to Know (November 2026)

Update date: November 4, 2026. Tencent has released the final version of Hy4, moving beyond the text-focused Hy4 Preview with visual understanding and computer-use capabilities. The headline is not simply that another large model arrived: Hy4 is being positioned for longer workflows in which an agent must interpret information, use tools, and work across graphical and command-line interfaces.

There is one important detail developers should not miss: Tencent says the final Hy4 model weights and system cards will be released in the coming days. The hosted model is available now, but the downloadable Hy4 Preview checkpoint is not automatically the same thing as the newly announced Hy4 release. If you want to test the new vision and computer-use capabilities today, use a hosted endpoint or Tencent's supported products rather than assuming the preview weights already contain them.

Who should pay attention?

• AI application developers evaluating a new model for image-aware assistants, tool calling, and long-running workflows.
• Agent builders working with browser, desktop, command-line, design, or coding tools.
• Security teams interested in how a model provider is limiting offensive cyber-training data while retaining defensive analysis and remediation tasks.
• Model infrastructure teams deciding whether to use a hosted API now or wait for weights, a system card, and serving recipes.

What Tencent announced on November 4

Tencent's announcement describes Hy4 as its latest-generation large language model. The new release adds visual understanding and computer use, alongside improvements to instruction following, long-context comprehension, and long-horizon agent tasks. Tencent says the model can process fine-grained images, analyse complex charts, and perform visual mathematical reasoning. It also says Hy4 can interact with command-line and graphical interfaces and received specialised training on professional software including Blender and CAD.

These are vendor-described capabilities. The announcement does not provide a public, independently reproduced evaluation of every workflow, and developers should test the exact tasks and tool harness they plan to deploy.

Hy4 versus Hy4 Preview

The official Hy4 Preview repository documents the preview architecture as a 770-billion-parameter mixture-of-experts model with 49 billion parameters activated per token and a one-million-token context configuration. Those are preview-checkpoint specifications. Do not automatically copy every architecture or runtime claim from that repository onto the final Hy4 release until Tencent publishes the final model card and weights.

Availability: hosted access first, final weights later

Tencent says Hy4 is available through Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, Yuanbao, WeCom's AI Assistant (Dayuan), and ima. Its announcement also names third-party API platforms including OpenRouter, Hermes, and OpenCode. The Tencent Cloud TokenHub model catalogue now lists Hy4 as a model with text generation, deep reasoning, and multimodal understanding.

OpenRouter lists the final model as tencent/hy4, separate from the older tencent/hy4-preview entry. Its model page currently lists text and image input, text output, tool calling, structured outputs, and a one-million-token context window. Provider routing, feature support, context limits, and prices can change, so verify the live model page before deploying.

Try the hosted Hy4 API with Python

The following example uses OpenRouter's OpenAI-compatible endpoint. It is a quick way to evaluate the hosted model without provisioning a multi-GPU server. It does not download model weights or make the model run locally.

1. Create an API key

Create an OpenRouter account and API key, then keep the key outside source control. Install the OpenAI Python SDK:

python -m venv .venv
source .venv/bin/activate
pip install openai

2. Set the key as an environment variable

export OPENROUTER_API_KEY="your_api_key"

On Windows PowerShell, use:

$env:OPENROUTER_API_KEY = "your_api_key"

3. Send a first request

import os
from openai import OpenAI

client = OpenAI(
    base_url="https://openrouter.ai/api/v1",
    api_key=os.environ["OPENROUTER_API_KEY"],
)

response = client.chat.completions.create(
    model="tencent/hy4",
    messages=[
        {
            "role": "system",
            "content": "You are a careful software engineering assistant."
        },
        {
            "role": "user",
            "content": "Explain how a Django request passes through middleware."
        },
    ],
    max_tokens=700,
)

print(response.choices[0].message.content)

Use the exact model identifier shown on the provider's current model page. A successful response confirms that your account can reach the hosted endpoint; it does not verify the model's performance for your workload.

4. Test image understanding separately

For a vision test, pass an image using the provider's supported image-content format. Use a non-sensitive image that you are authorised to upload:

response = client.chat.completions.create(
    model="tencent/hy4",
    messages=[
        {
            "role": "user",
            "content": [
                {"type": "text", "text": "Summarise this chart and identify any unclear labels."},
                {
                    "type": "image_url",
                    "image_url": {"url": "https://example.com/sample-chart.png"}
                },
            ],
        }
    ],
    max_tokens=500,
)

print(response.choices[0].message.content)

Replace the sample URL with an image that your provider can access. If your application sends private images, check the provider's data handling and retention terms first. Do not treat a successful image response as proof that every chart, diagram, or screenshot will be interpreted correctly.

Computer use is not the same as ordinary tool calling

Tencent says Hy4 can interact with graphical and command-line interfaces. That does not mean a plain chat-completions request automatically gains access to your desktop. An agent runtime must expose the relevant tools, capture screen state, translate actions into supported operations, and return observations to the model.

For a first evaluation, keep the model separate from privileged production systems:

1. Use a disposable virtual machine or isolated test account.
2. Start with read-only tasks such as identifying a button, summarising a visible dashboard, or locating a setting.
3. Expose only a small allowlist of actions; do not give the agent an unrestricted shell or administrator account.
4. Require human confirmation before sending messages, deleting files, changing infrastructure, submitting forms, or spending money.
5. Record screenshots, tool arguments, tool results, and the final answer so failures can be reproduced.
6. Test prompt injection from web pages, documents, images, and terminal output before allowing broader access.

Function calling and computer use are related but different. Function calling asks the model to select a structured tool. Computer use adds an observation-and-action loop over a graphical or command-line environment. Treat the latter as a higher-risk capability.

What the launch says about cybersecurity and biosafety

Tencent says Hy4 was trained with defensive cybersecurity data, prioritising proof-of-concept development and vulnerability remediation. The company says it excluded offensive exploitation material, while retaining reverse engineering and penetration-testing material for analysis. Tencent's Xuanwu Lab and Keen Security Lab conducted internal assessments; Tencent reports no significant concerns in the offensive-capability areas they evaluated.

Reuters independently reported the release and Tencent's stated restrictions, but the detailed safety conclusions remain company-reported. The public system card and full assessment report were not available in the launch announcement. A provider's training choices and internal tests are useful information, not a guarantee that a deployed model cannot produce harmful output or be misused.

On biosafety, Tencent says internal expert reviews found strong foundational biological knowledge but more limited open-ended experimental-design capability than frontier models. This is also a vendor-reported assessment, not an independent certification.

How to interpret the performance claims

Tencent reported several internal and partner evaluations:

• In a financial-task blind evaluation involving 213 professionals and 579 tasks, Tencent reported an average score of 3.04 for Hy4, compared with 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 for Kimi K3.1 Pro and GLM-5.3.
• Tencent reported a WorkBuddy benchmark average of 83.18 and a 96.15% pass rate, with first place in nine of twelve evaluated scenarios in the mainland China market.
• Tencent reported an internal search-evaluation hallucination rate change from 12.95% for Hy4 Preview to 7.2% for Hy4, and unsupported-assumption errors falling from 23.1% to 5.6%.

These figures are not independent benchmark results. They use particular task sets, scoring methods, and evaluation environments. The financial evaluation included external professionals, but it was still organised by Tencent and Tencent Financial Technology. Treat the numbers as reasons to run your own evaluation, not as proof that Hy4 will outperform another model in your application.

Pricing and access windows

Tencent's announcement gives approximate standard API prices of US$0.90 per million input tokens, US$2.70 per million output tokens, and US$0.045 per million cached input tokens. OpenRouter currently displays provider-specific pricing and a 50% promotional rate on its Hy4 page. These figures can differ because of provider routing, currency conversion, caching, and temporary offers; check the live rate card before estimating production spend.

Tencent also announced free WorkBuddy and CodeBuddy access from November 4 through November 15, 2026, and a 50% TokenHub discount through December 31, 2026. Availability can depend on account, region, and product conditions. Confirm the offer in the relevant account rather than assuming it applies universally.

Limitations and what is still unknown

• Final weights are pending: Tencent says the final weights and system cards will follow in the coming days. The preview checkpoint is available, but it is not a substitute for the final release.
• Independent testing is limited: the launch-day performance and safety figures are mainly vendor or partner evaluations.
• Computer-use integration needs a harness: developers must supply tools, permissions, observation handling, and approval gates.
• Hosted API means provider dependence: latency, availability, data handling, and price depend on the selected provider and route.
• Open-weight status needs precise wording: Tencent has announced an open-source release plan, but the final Hy4 weights and system card were not yet published in the launch announcement. The already downloadable Hy4 Preview is a distinct checkpoint.

Who benefits most?

Hy4 is worth evaluating if you are building image-aware assistants, coding or office agents, chart-analysis workflows, or controlled computer-use systems and want to test a new hosted model quickly. Teams that require self-hosting, reproducible model files, or a full safety review should wait for the final weights, license details, and system card before making a deployment decision.

Frequently asked questions

Is Tencent Hy4 open source?

Tencent announced that it plans to release the final model weights and system cards in the coming days. At the time of this update, the final release is accessible through hosted services, while the older Hy4 Preview weights remain separately available. Avoid describing the final checkpoint as downloadable until the release files are actually published.

Can I run the new Hy4 locally with Ollama or LM Studio?

Do not assume so. The final weights have not yet been published in the announcement, and local runtime compatibility depends on the eventual model format, architecture support, quantisation, and hardware requirements. The preview repository's vLLM and SGLang recipes apply to the preview checkpoint, not automatically to the final Hy4 release.

Does Hy4 support images?

Yes. Tencent says the November 4 release introduces visual understanding. OpenRouter also lists image input for its tencent/hy4 endpoint. Confirm the exact request format and limits with the provider you use.

Does Hy4 have computer-use capabilities?

Tencent says Hy4 can work with graphical and command-line interfaces. To use that capability in an application, you still need a compatible agent harness and carefully permissioned tools. Do not grant unrestricted desktop or shell access.

Is Hy4 safer for cybersecurity work?

Tencent reports defensive-only training priorities and internal assessments. Those claims are not an independent guarantee. Evaluate the model with your own authorised defensive test suite, restrict tools, and keep human review for high-impact actions.

Sources and further reading

• Tencent's November 4, 2026 Hy4 announcement distributed by PR Newswire — capabilities, evaluation claims, safeguards, pricing, and planned weight release.
• Reuters, November 4, 2026 — independent reporting on the release and Tencent's stated cyber and biosafety restrictions.
• Tencent Cloud TokenHub — current hosted model catalogue.
• OpenRouter: Tencent Hy4 — current model identifier, API integration, modalities, and provider pricing.
• Tencent Hy4 Preview repository — preview model details and its separate serving instructions.

Related GyanAangan guides

• Akamai Inference and Agent Sandbox: Open-Weight APIs Meet WebAssembly Isolation — compare hosted inference and isolated agent execution.
• Gemini 4 Carbon: Cybersecurity Benchmarks and Developer Access — another recent model release with restricted cyber access.
• Mistral Large 4: What Developers Need to Know About Open Weights — a useful contrast for teams comparing hosted access with downloadable weights.
• Build a Local LLM Gateway with LiteLLM — useful when routing between hosted and local model endpoints.
• Secure Multi-Tenant MCP Servers with OAuth and Tenant Isolation — relevant background for permissioned tools and agent integrations.

Editorial note: This article was prepared from the sources linked above and updated on November 4, 2026. No independent Hy4 benchmark, live API test, or local deployment test was performed for this article. Vendor claims are labelled as such, and final Hy4 weights/system-card availability should be rechecked before implementation.