# Hy4

> Source: https://aiwiki.ai/wiki/hy4
> Updated: 2026-11-05
> Categories: AI Models, Chinese AI, Large Language Models, Multimodal AI
> License: CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/) - attribute to "AI Wiki (aiwiki.ai)"
> Cite as: AI Wiki. "Hy4." aiwiki.ai, 5 Nov 2026. https://aiwiki.ai/wiki/hy4
> From AI Wiki (https://aiwiki.ai), the free encyclopedia of artificial intelligence. Reuse freely with attribution.

**Hy4** is a [large language model](https://aiwiki.ai/wiki/large_language_model) in [Tencent's](https://aiwiki.ai/wiki/tencent) [Hy (Hunyuan)](https://aiwiki.ai/wiki/hunyuan) family, released through hosted services on November 4, 2026. It adds image understanding and [computer use](https://aiwiki.ai/wiki/computer_use) to the capabilities of [Hy4 Preview](https://aiwiki.ai/wiki/hy4_preview), with further pre-training and post-training for coding, research and workplace tasks.[1][2]

The hosted release and the planned weight release are separate. Tencent's launch announcement says that model weights and a system card will follow. Its description of Hy4 as an open model therefore states the intended release approach; the announcement does not establish that the final weights were already downloadable on launch day.[1][2]

## Release and API configuration

Tencent lists access through Tencent Cloud and OpenRouter, alongside integration into CodeBuddy, WorkBuddy, Yuanbao and ima. Some capabilities described in the launch depend on those products and their connected tools, rather than on a text request to the model alone.[1]

OpenRouter identifies the final model as `tencent/hy4`, separately from `tencent/hy4-preview`. Its launch-day model registry provides the following serving configuration.[3]

| Property | OpenRouter listing, November 4, 2026 |
| --- | --- |
| Model identifier | `tencent/hy4` |
| Versioned identifier | `tencent/hy4-20261104` |
| Inputs | Text and images |
| Output | Text |
| Context limit | 1,048,576 tokens |
| Provider maximum completion | 65,536 tokens |
| Reasoning efforts | `none`, `low`, `high` |
| Default reasoning effort | `high` |
| Other listed request features | Tool calling and structured outputs |

These values describe the listed service configuration. They should not be treated as a self-hosting specification for the forthcoming weights.[1][3]

## Vision and computer use

Hy4 introduces visual perception within this model line. Tencent describes fine-grained image inspection, chart interpretation and mathematical visual reasoning. Its demonstrations include locating a requested component in a large aerial image. The demonstration is an example of the supplied task, rather than a measure of general image-grounding accuracy.[1]

The computer-use capability combines command-line and graphical-interface interaction. Tencent says it trained on professional software including Blender and CAD applications, as well as workflows that span multiple applications. This connects visual observations to reasoning and tool calls over a sequence of steps.[1][2]

The model's listed API output is text. The software actions, editable documents and rendered scenes in product demonstrations also involve the surrounding application, its tools and execution environment.[1][3]

## Coding and game development

Tencent describes training and evaluation work with developers from its Interactive Entertainment Group. The stated targets include implementing features, maintaining gameplay state, following game logic and handling asset pipelines. The release also describes adaptation for Unreal Engine, Unity and Godot, together with multiplayer-development tasks.[1]

For frontend work, the stated focus includes business logic, state management, interactions and full-stack integration. Tencent reports improvements over Hy4 Preview of 19.1% on GameDev and 9.8% on GameLogic. These figures belong to Tencent's evaluations; they do not measure every project or establish that a generated game is ready for release without testing.[1][2]

For the surrounding development workflow, see [Developing Games with AI](https://aiwiki.ai/wiki/developing_games_with_ai).

## Financial and office tasks

Tencent says it built thousands of financial workspaces and tens of thousands of tasks spanning fund research, equity analysis, quantitative strategies, asset allocation, corporate finance and tax accounting. It also describes search improvements intended to retrieve timely information, combine sources and check factual consistency.[1]

The company and Tencent Financial Technology conducted a blind evaluation with 213 financial professionals, including 153 external specialists, across 579 financial tasks. The launch materials report these overall performance scores:[1][2]

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

These are scores from the described evaluation, not percentages or independent measures of investment performance. Separately, RabyteTech's launch-page statement reports a score of 61.63 on 101 investment-research tasks in its IRAB benchmark, 12.03 points above Hy4 Preview.[1]

Tencent also describes editing Word, Excel and PowerPoint documents with people. WorkBuddy's WeCom integration supplies access to work context such as messages, email, documents, calendars and meeting records. Team collaboration in a new Projects version is described as forthcoming in the launch announcement.[1]

## Scientific research

TenureTrail is Tencent's internal research benchmark containing 123 tasks across seven disciplines. Tencent reports that Hy4 ranks first among the open-weight models in its comparison, including tasks in mathematical construction, physical modeling, earth science, systems engineering and socioeconomic decision-making.[1]

The release also describes work with the [Hyra 2.0](https://aiwiki.ai/wiki/hyra_2_0) research harness on pre-training infrastructure and model development. Tencent announces HyCell for single-cell representation and HyPDE for physical-field modeling as reusable research outputs co-developed with the system.[1][2]

These reports concern particular research workflows and outputs. They do not establish an unrestricted autonomous process for improving Hy4 itself. Tencent also acknowledges a remaining gap with leading proprietary models on difficult internal tasks in mathematics, extended coding, model improvement and advanced visual question answering.[1]

## Launch pricing

The following figures distinguish Tencent's launch-page prices from the discounted prices in OpenRouter's model registry. They are a dated snapshot, rather than a promise of permanent rates.[1][3][4]

| Per million tokens | Tencent launch page | OpenRouter registry, November 4, 2026 |
| --- | ---: | ---: |
| Uncached input | $0.834 | $0.417 |
| Cached input | $0.042 | $0.021 |
| Output | $2.501 | $1.2505 |

OpenRouter labels its listed Tencent Cloud endpoint 50% off. Tencent's press release separately announces a Tencent Cloud TokenHub promotion at half the standard price through December 31, 2026. That release does not establish the end date of every reseller's promotion.[2][4]

## Safety evaluations and weight release

Tencent says it used defensive cybersecurity training material and red-team assessments involving Xuanwu Lab and Keen Security Lab. Its launch page reports a score of 30.4/100 on its Hy-BioExperimentBench assessment of open-ended biological experimental design, below the company's stated high-risk threshold. These are developer-reported evaluations, not proof that the model cannot produce harmful outputs.[1]

Tencent's November 4 press release says the cybersecurity corpus prioritized defensive proof-of-concept development and vulnerability remediation, while excluding offensive exploitation material.[2] The launch page and press release give different accounts of reverse engineering and penetration testing: the launch page groups them with excluded offensive content, while the press release says they were included for analysis purposes. The published descriptions do not resolve this difference.[1][2]

The company plans to publish its system card alongside the weights and says it is working with the [vLLM](https://aiwiki.ai/wiki/vllm) and [SGLang](https://aiwiki.ai/wiki/sglang) communities on ecosystem support. The final model's architecture, licensing terms and self-hosting requirements should be taken from those release artifacts when published; the corresponding details of Hy4 Preview do not automatically apply to the final release.[1]

Hy4 Flash is also described as a future release targeting faster inference and lower costs. It is separate from the Hy4 model released on November 4.[1][2]

## References

1. Tencent Hy. [Introducing Hy4](https://hy.tencent.ai/research/hy4). Official release article, November 4, 2026; rendered page checked November 4, 2026.
2. Tencent. [Tencent Launches Hy4: Joining the Ranks of the World's Leading Open-Source AI Models](https://en.prnasia.com/releases/apac/tencent-launches-hy4-joining-the-ranks-of-the-world-s-leading-open-source-ai-models-554689.shtml). Company-issued release distributed by PR Newswire, November 4, 2026.
3. OpenRouter. [Models API](https://openrouter.ai/api/v1/models), `tencent/hy4` entry. Serving limits, modalities, reasoning settings and exact per-token prices, retrieved November 4, 2026.
4. OpenRouter. [Tencent: Hy4](https://openrouter.ai/tencent/hy4). Model and provider listing, including the 50% discount label, checked November 4, 2026.