Tencent Hy4 Launched: Vision, Computer Use and Open Weights Coming (Benchmarks and Pricing)

Tencent has launched Hy4, the newest version of its Hunyuan-family large language model, and says it will publish the open weights and system cards “in the coming days”. Hy4 adds image understanding and computer use (operating command-line and desktop apps, with special training on Blender and CAD), cuts Tencent’s measured hallucination rate from 12.95% to 7.2%, and is already live on OpenRouter at about $0.90 per million input tokens. Here is what is confirmed, what the benchmarks actually show, and what to watch for when the weights land.

What Is New in Hy4

Hy4 replaces Hy4 Preview. Tencent lists four main upgrades:

• Visual understanding is new to this generation: fine-grained image detail, complex chart reading and mathematical reasoning over images.
• Computer use is also new. The model can drive command-line and graphical interfaces, and Tencent says it was specially trained on Blender and CAD software to plan and carry out long, multi-step tasks.
• Better instruction following and long-context reading, plus more efficient reasoning.
• Long-horizon agent work, which is the main focus of the release.

The Blender angle is interesting for our readers. We have covered the best AI for Blender animation and how to drive Blender from a local LLM through MCP. A model trained directly on Blender workflows, with open weights on the way, could become a strong local option for that use case.

Benchmarks: What Tencent Published

Most numbers in the announcement are Tencent’s own internal or partner tests, not public leaderboards. Here are the concrete ones:

Tencent did not publish computer-use scores (such as OSWorld) or the usual public coding and reasoning benchmarks. Third-party tracker Opper lists Hy4 with an “intelligence” score of 35, well below frontier closed models like Gemini 4 Carbon (60) and GPT-6.1 Terra (51) on the same site. So the honest read is: Hy4 looks like a strong, cheap agent model for business and finance tasks, but not a frontier challenger on raw reasoning.

Independent Scores (Artificial Analysis)

Artificial Analysis has already tested Hy4 through the API, and its numbers give a more neutral picture than Tencent’s own tables. Opper republishes these scores on its Hy4 page:

The pattern is clear: Hy4 is much stronger at agent and automation work (Agentic Index 51.5, AutomationBench 66%) than at hard coding and science reasoning. For reference, Gemini 4 Carbon scores 60 on the same Intelligence Index and 65% on Terminal-Bench 4.0, so Hy4 sits a full tier below the frontier, but at a fraction of the price.

Pricing and Availability

• API: about $0.90 per million input tokens, $2.70 per million output tokens and $0.045 per million cached input tokens, the same as Hy4 Preview.
• Discount: Tencent Cloud TokenHub offers 50% off standard pricing until 31 December 2026.
• Free: Tencent’s WorkBuddy and CodeBuddy tools are free from 4 to 15 November.
• Third-party: OpenRouter, Hermes and OpenCode, so you can plug it into most coding agents today.
• Apps: Yuanbao, WeCom’s assistant and ima now, with QQ, WeRead and Tencent News integrations under way.

For comparison, that input price is in the same budget tier as DeepSeek’s Flash models. See our DeepSeek vs Claude breakdown for how the cheap Chinese models stack up against Anthropic’s line-up.

How to Use Hy4 Today (Step by Step)

Option 1: free chat, no setup. Tencent’s Yuanbao app and the ima assistant already run Hy4. They are aimed at Chinese users, so the interface is mostly Chinese.

Option 2: free coding tools until 15 November. Tencent’s CodeBuddy (coding assistant) and WorkBuddy (office agent) are free from 4 to 15 November, which is the cheapest way to test Hy4’s agent skills.

Option 3: OpenRouter API (easiest for most developers).

1. Create an account at openrouter.ai and add a small amount of credit.
2. Create an API key under Keys.
3. Search the Models page for “Hy4” and copy its exact model ID.
4. Call it with any OpenAI-compatible client by setting the base URL to https://openrouter.ai/api/v1:

from openai import OpenAI

client = OpenAI(base_url="https://openrouter.ai/api/v1",
                api_key="YOUR_OPENROUTER_KEY")

reply = client.chat.completions.create(
    model="PASTE_HY4_MODEL_ID_FROM_OPENROUTER",
    messages=[{"role": "user", "content": "Summarise this chart and list three risks."}],
)
print(reply.choices[0].message.content)

Option 4: inside a coding agent. OpenCode lists Hy4 directly. In Cline, Roo Code or Continue, choose OpenRouter as the provider, paste your key and pick the Hy4 model. If you already use local models in these tools, our Cline + Blender MCP guide shows the same provider setup, and you can point Blender MCP at Hy4 to test its Blender training.

Option 5: Tencent Cloud TokenHub for businesses that want the 50% launch discount and direct billing from Tencent.

Hy4 vs DeepSeek, Kimi, Gemini and Claude

A dash means we did not find a comparable published score. On Tencent’s own finance blind test, Hy4 (3.04) narrowly beat DeepSeek-V4.1-Flash (2.97) and clearly beat Kimi K3.1 Pro and GLM-5.3 (both 2.68). Price-wise it sits between DeepSeek Flash and the $2/$10 tier. For a fuller look at DeepSeek and Claude pricing and benchmarks, see our DeepSeek vs Claude comparison, and for running the open Kimi model yourself, our Kimi K3 VRAM guide.

Can You Run Hy4 Locally?

Not yet. The weights are not out, and Tencent has not said how big the model is. When they appear, check three things before you get excited: the parameter count and active parameters (if it is a big mixture-of-experts model, you will need lots of system RAM, as with Kimi K3); the licence (Tencent’s earlier Hunyuan licences had regional limits); and whether llama.cpp and Ollama support the architecture on day one. We will publish a run-locally guide with VRAM tables as soon as the files are on Hugging Face. Until then, our best Ollama models list covers what you can run today.

Safety Notes

Tencent says Hy4 was trained on defensive security data (proof-of-concept work, vulnerability fixes, reverse engineering for analysis) with offensive exploitation material removed, and that its Xuanwu and Keen security labs found no significant concerns. On biology, Tencent describes strong base knowledge but weaker open-ended experiment design than frontier models.

What We Still Do Not Know

• Size: total and active parameter counts, and whether Hy4 is a mixture-of-experts model.
• Context window and max output: not listed by Tencent or the trackers.
• Licence: whether the open weights allow commercial use and in which regions.
• Computer-use scores: no OSWorld or similar results, despite computer use being a headline feature.
• Hy4 Flash: size, price and release date.

We will update this article when the weights and model card are published.

My Take

The open-weights promise is what makes Hy4 worth watching. A cheap API model with vision and computer use is nice; an open one trained on Blender and CAD would be a gift to the local-agent crowd. Until the weights and a model card land, treat the benchmark table as marketing from Tencent, and watch for independent scores.

Frequently Asked Questions

Is Tencent Hy4 open source?

Tencent says it will release the model weights and system cards in the coming days. The licence has not been published yet.

How much does the Hy4 API cost?

About $0.90 per million input tokens and $2.70 per million output tokens, with 50% off on Tencent Cloud TokenHub until the end of 2026.

Can Hy4 control my computer?

Hy4 supports computer use for command-line and graphical apps, and was trained on Blender and CAD tasks. No computer-use benchmark scores were published.

What is Hy4 Flash?

A smaller, faster Hy4 model that Tencent says is coming soon. No specs yet.

How good is Tencent Hy4 compared with GPT and Gemini?

On Artificial Analysis it scores 34.6 on the Intelligence Index and ranks 17th of 507 models, below frontier models such as Gemini 4 Carbon (about 60) but strong for its price, especially on agent tasks (Agentic Index 51.5).

How do I use Hy4 in English?

Use it through OpenRouter or OpenCode, or plug OpenRouter into Cline, Roo Code or Continue. Tencent’s own consumer apps are mostly in Chinese.

What is Hy4’s context window?

Tencent has not published it yet. Check the model card when the open weights are released.

Sources

• Tencent: Tencent launches Hy4
• Opper: Tencent Hy4 model page
• NoCode MBA: AI models released in November 2026