Tencent Hy4: Vision, Computer Use, API Pricing and Setup

Last updated: November 4, 2026

Tencent Hy4, released November 4, 2026, is the full version of Tencent’s open model line. It adds image understanding and computer use for the first time, improves long-context and agent work over Hy4 Preview, keeps the same API price and is live on OpenRouter as tencent/hy4 with a 1M-token context window.

Contents

What’s new in Hy4
Hy4 vs Hy4 Preview
Where to use Hy4 and what it costs
How to call Hy4 through the API
Benchmarks Tencent reported
Workflows and prompt patterns
Limitations
Frequently asked questions

Tencent announced the release in its Hy4 launch post, about ten weeks after Hy4 Preview shipped on August 28, 2026. This guide covers what changed, how to reach the model and what the published numbers do and do not show.

What’s new in Hy4

Feature	What changed
Visual understanding	Reads images for the first time, including fine-grained image work, complex charts and math problems shown visually
Computer use	Can operate command-line and graphical interfaces; trained on professional software such as Blender and CAD tools
Agent and long context	Better instruction following, long-context comprehension and long-horizon agent tasks
Search accuracy	Targeted search tuning; internal hallucination rate fell from 12.95% to 7.2%
Game development	Integrations with Unreal Engine, Unity and Godot; GameDev up 19.1% and GameLogic up 9.8% over the preview
Research models	Tencent open-sourced HyCell (single-cell representations) and HyPDE (physics fields), built with Hy4 and the Hyra 2.0 harness

All figures in this table come from Tencent’s own internal evaluations in the launch post.

Hy4 vs Hy4 Preview

Aspect	Hy4 Preview	Hy4
Release date	August 28, 2026	November 4, 2026
Input	Text only	Text and images
Computer use	No	Yes, CLI and GUI
Context window	1,048,576 tokens	1,048,576 tokens
Max output on OpenRouter	64,000 tokens	65,536 tokens
Open weights	Published on Hugging Face	Planned in the coming days
List price per 1M tokens (in / out / cached)	US$0.834 / $2.501 / $0.042 on OpenRouter	About US$0.90 / $2.70 / $0.045 per Tencent; $0.834 / $2.501 / $0.042 list on OpenRouter

The Hy4 Preview model card describes a mixture-of-experts model with 770 billion total and 49 billion active parameters. Tencent has not yet published a separate specification for the final Hy4, so do not assume the size is unchanged until the system card arrives. The preview card also listed known issues, including spending longer than needed on reasoning and over-verifying its own work.

Where to use Hy4 and what it costs

Tencent apps: WorkBuddy, CodeBuddy, Yuanbao, the WeCom AI Assistant (Dayuan) and ima.
APIs: Tencent Cloud TokenHub, plus third-party platforms including OpenRouter, Hermes and OpenCode.
Price: about US$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens, unchanged from the preview.
Launch offers: WorkBuddy and CodeBuddy are free from November 4 to 15, and TokenHub is 50% off through December 31, 2026.

On OpenRouter, the tencent/hy4 listing shows a single provider, Tencent Cloud, at a discounted US$0.417 input, US$1.251 output and US$0.021 cache read per million tokens. It is not currently available through OpenRouter’s EU in-region routing.

How to call Hy4 through the API

OpenRouter is the quickest route outside China because it uses an OpenAI-compatible endpoint.

Create an API key in your OpenRouter dashboard and export it as OPENROUTER_API_KEY.
Point any OpenAI-compatible SDK at https://openrouter.ai/api/v1, or send a POST to /api/v1/chat/completions.
Set "model": "tencent/hy4" and send your messages. Images go in as standard image content parts.
Turn on thinking with the reasoning parameter or reasoning_effort, and pass back the full reasoning_details on later turns so the model can continue its reasoning.
For agents, send tools and tool_choice. For strict output, use a JSON schema in response_format.
Cap long answers with max_tokens; the listing allows up to 65,536 completion tokens.

OpenRouter also exposes the model through Responses API and Anthropic Messages formats, which is why agent tools such as Claude Code already send it traffic, according to the listing’s app rankings.

Benchmarks Tencent reported

In a blind finance evaluation run by Tencent with Tencent Financial Technology, 213 professionals scored 579 real tasks. Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. See our DeepSeek V4.1 Flash guide and Kimi K3.1 Pro guide for those rivals.

WorkBuddy: average score 83.18 and a 96.15% pass rate, first in nine of twelve scenarios in mainland China.
Investment research: 12.03 percentage points better than Hy4 Preview on Rabyte’s internal iRaB benchmark.
Unsupported assumptions: errors fell from 23.1% to 5.6%.
Science: validated on TenureTrail, 123 executable tasks across seven disciplines.

These are vendor-run results. Treat them as Tencent’s claims until independent tests are published.

Workflows and prompt patterns

Chart and report analysis. Attach a chart image and ask: “Extract every data series into a table, then list three trends and any figures that look inconsistent.”
Long-context code review. Load a large repository slice into the 1M-token window and ask for a dependency map before asking for changes.
Cheap coding agent. Use Hy4 as the model in a terminal agent such as OpenCode, which Tencent names as a launch platform.
Game prototypes. Name the engine up front (“Godot 4, GDScript, two-player local multiplayer”) so the model uses the engine integrations Tencent trained on.
Fact-checked research. Ask it to cite each source and flag any claim it could not verify, which plays to the search tuning.

For other large open models released this autumn, compare our LongCat-2.5 guide, MiniMax M3.1 Flash guide and Muse Spark Open guide.

Limitations

Final Hy4 weights and the system card are not published yet, so you cannot self-host the release version today.
Tencent says Hy4 Flash is coming soon but has not given a date or price.
On OpenRouter only Tencent Cloud serves the model, so there is no failover to other providers.
Tencent says the model’s open-ended biology experiment design is weaker than frontier models, and offensive exploit material was excluded from training.
The 50% discounts are temporary; budget at the full list price after December 31, 2026.

Frequently asked questions

What is Tencent Hy4?

Tencent Hy4 is the official release of Tencent’s latest large language model, announced on November 4, 2026. It follows Hy4 Preview from August and adds image understanding and computer use for the first time, along with better instruction following, long-context handling and long-running agent tasks.

Is Tencent Hy4 open source?

Not yet. Tencent says it plans to release the Hy4 model weights and system cards in the coming days. The earlier Hy4 Preview weights are already published on Hugging Face as tencent/Hy4-preview.

How much does the Hy4 API cost?

Tencent lists about US$0.90 per million input tokens, US$2.70 per million output tokens and US$0.045 per million cached input tokens, the same as Hy4 Preview. Tencent Cloud TokenHub offers 50% off through December 31, 2026, and OpenRouter lists the model at the discounted rate.

What is the Hy4 context window?

OpenRouter lists a context window of 1,048,576 tokens for tencent/hy4, with up to 65,536 completion tokens per response. The model accepts text and images and returns text.

Where can I use Hy4 for free?

Tencent says WorkBuddy and CodeBuddy are free from November 4 to 15, 2026. Hy4 also runs in Yuanbao, ima and the WeCom AI Assistant.

What is Hy4 Flash?

Hy4 Flash is a lower-cost version that Tencent says is scheduled for release soon. Tencent has not published its pricing or specifications yet.

Sources

Image: Tencent Hy logo from the Tencent-Hunyuan GitHub repository, Apache License 2.0, via Wikimedia Commons. Tencent is a trademark of Tencent Holdings Ltd. FreeAnswers is not affiliated with or endorsed by Tencent.

Tencent Hy4: Power-User Guide to Vision, Computer Use, API and Pricing