techcoff.com

What Is Tencent Hy4? The 770B AI That Sees Your Screen and Clicks for You

What is Tencent Hy4 — 770B AI model with computer use explained

A 770-billion-parameter AI that can look at your screen, read a chart, click the right button, and finish a multi-step task across your apps — without you touching the mouse. That is the promise Tencent made on November 4, 2026, when it released Hy4, the latest model in its Hunyuan (Hy) family. The headline is not a bigger benchmark score. It is a new kind of ability: computer use.

So what is Tencent Hy4, really? Is it open source, how much does it cost, and should you care if you already use GPT or Claude?

What Is Tencent Hy4?

Tencent Hy4 is Tencent’s flagship large language model, released on November 4, 2026, as the successor to Hy3 and the full version of the Hy4 preview from August 28, 2026. It is a mixture-of-experts (MoE) model with 770 billion total parameters and a 1-million-token context window — roughly 1,048,576 tokens.

The preview was text-only. The full Hy4 adds the two capabilities that define it: visual understanding and computer use — it reads images and charts, then acts on what it sees through the command line and interfaces.

The key takeaway: Hy4 is built to be an agent that does the work, not just a chatbot that answers questions.

What is Tencent Hy4 — 770B AI model with computer use explained

Tencent Hy4 Computer Use: What It Can Actually Do

Computer use is the feature that separates Hy4 from the August preview and from most rival models. It reads on-screen text, buttons, and menus, so it can navigate software it was never hard-coded for. It clicks UI elements, fills forms, and moves between windows — and it also works through the command line. Most importantly, it chains long tasks across multiple applications instead of stopping after a single prompt. Tencent trained it on professional software, including Blender and CAD tools, plus game engines — Unreal Engine, Unity, and Godot.

The key takeaway: if your work lives inside professional software, Hy4 was trained specifically for your world.

A Real-World Example: The Friday Report That Builds Itself

Picture Friday afternoon: the sales data sits in a spreadsheet, the charts belong in a slide deck, and a summary has to reach your team by email — normally 45 minutes of copying and pasting. With Hy4, you describe the finished report once. It reads the spreadsheet, builds the charts, lays out the deck, and drafts the email, clicking through each app the way you would. That is the benefit Tencent is selling: not a smarter answer, but a finished task.

Tencent Hy4 Specs at a Glance

The most detailed public architecture numbers come from the Hy4 preview’s model card: 770B total parameters (MoE), 49B active per token, 78 layers with 256 routed experts plus 1 shared expert, and a 1M-token context window. New in the full model: vision and computer use (CLI + GUI).

The key takeaway: only 49B of the 770B parameters activate per token.

Hy4 vs Hy4 Preview vs Hy4 Flash

Here is how the three tiers compare, using only verified figures:

FeatureHy4 (full)Hy4 PreviewHy4 Flash
ReleasedNov 4, 2026Aug 28, 2026Nov 16, 2026
Total parameters770B770B224B
Active per token49B49B12B
Vision + computer useYesNoYes
Input price / 1M tokens$0.834$0.834$0.07
Output price / 1M tokens$2.501$2.501$0.281

Hy4 Flash: The Cheap, Fast Version for Office Work

Hy4 Flash arrived on November 16, 2026 — and it may matter more to most users than the flagship. It is a 224-billion-parameter model with only 12 billion active parameters per token, yet it keeps the 1-million-token multimodal context window. Tencent positions Flash for everyday work: research, slides, spreadsheets, and documents. On Tencent’s internal WorkIndex benchmark, Flash cost about one-twentieth as much per task as the full Hy4 model — $0.07 per million input tokens and $0.281 per million output tokens, roughly 1/12 of full Hy4 on input. Free in WorkBuddy through November 30, 2026.

The key takeaway: to try the Hy4 family without the flagship price tag, start with Flash.

Tencent Hy4 Pricing: What It Costs

  • OpenRouter (model ID: tencent/hy4): $0.834 per 1M input tokens and $2.501 per 1M output tokens, with a promotional rate near $0.42/$1.25.
  • Tencent Cloud TokenHub: RMB 6 per 1M input tokens and RMB 18 per 1M output tokens, with 50% off through December 31, 2026.
  • Hy4 Flash: $0.07 per 1M input tokens and $0.281 per 1M output tokens, with 50% off on TokenHub through November 30, 2026.

The key takeaway: Flash is the budget door into the Hy4 family; the full model is priced like a flagship.

Hy4 Benchmarks: The Honest Picture

Nearly all Hy4 benchmarks are Tencent-reported, with no independent verification for the full model yet. And Tencent’s own numbers are modest: the company says Hy4 does not rank first on any of its 12 reported benchmarks, placing roughly 3rd–4th out of 7 models on most of them. Tencent openly admits a gap to Claude and GPT on some tasks. The one independent data point so far: Artificial Analysis lists Hy4 at 75.8% pass@1 on CWE-Bench-AA (defensive security tasks) as of November 13, 2026.

The key takeaway: Hy4’s pitch is capability — vision plus computer use — not a leaderboard crown.

Hy4 vs GPT-6.2: How Does It Compare?

GPT-6.2 (codenamed Astra Sol) is OpenAI’s flagship general-purpose model — our what is GPT-6.2 guide covers it in detail. It leads on broad reasoning and has the deepest US app ecosystem. Hy4’s edge is narrower but real: computer use trained on professional creative software — Blender, CAD, Unreal Engine, Unity, and Godot — plus aggressive pricing, especially through Hy4 Flash. Our ChatGPT Remix guide shows the same shift from the OpenAI side: models that make things, not just answer questions.

The key takeaway: GPT-6.2 wins on general smarts and ecosystem; Hy4 wins on hands-on computer work.

How to Use Tencent Hy4

  1. OpenRouter: the fastest route for most developers — use model ID tencent/hy4 (or tencent/hy4-flash for Flash) through the OpenRouter API or playground.
  2. Tencent Cloud TokenHub: Tencent’s own API platform, 50% off through December 31, 2026.
  3. CodeBuddy, WorkBuddy, Yuanbao, and WeCom: Tencent’s coding, office, and workplace apps.

One caution: the model clicks real buttons and runs real commands, so give it a sandbox or a test machine first, never your main work computer with live data. Exploring AI tools for your business? See our services or get in touch.

Is Tencent Hy4 Really Open Source?

Tencent calls Hy4 open source and says it plans to release the weights and a system card. But as of November 16–19, 2026, the full model’s weights had NOT been publicly released. Only the Hy4 preview weights from August 28 are downloadable under Apache 2.0. Until the weights land, Hy4 is an API model — and self-hosting would be a data-center job anyway, since the preview’s weights need roughly 924 GB quantized.

The takeaway: “open source” is currently a promise, not a download.

Tencent Hy4 Limits: What It Cannot Do Yet

The first limit is the benchmarks. Nearly all Hy4 benchmarks are Tencent-reported, with no independent verification for the full model. Tencent itself says Hy4 does not rank first on any of its 12 reported benchmarks, which is honest but also telling. A model that cannot top the charts it chose itself is not the strongest reasoner on the market. Tencent Hy4 wins on what it can do, not on how it scores.

The second limit is the open source promise. Tencent calls Hy4 open source, but the full model weights had not been publicly released as of mid-November 2026. Only the Hy4 preview weights from August 28 are downloadable under Apache 2.0. Until the full weights land, you cannot self-host the model, inspect it, or run it offline. For teams that need real open source, Tencent Hy4 is a promise to wait for, not a download to use today.

The third limit is size. A 770-billion-parameter model is not something you run on a laptop or a small server. The preview weights need roughly 924 GB quantized, so self-hosting the full model would be a data-center job even if the weights were public. For most users, Tencent Hy4 through the API is the only realistic route, through OpenRouter or Tencent Cloud TokenHub. That is fine for trying the model, but it means your data leaves your network on every request.

The fourth limit is computer use itself. A model that clicks real buttons and runs real commands can also click the wrong button and run the wrong command. Tencent trained Hy4 on professional software, but a misclick in a live spreadsheet, a CAD file, or a production database is still a misclick. That is why the advice in this post matters: give the model a sandbox or a test machine, not your real accounts. Tencent Hy4 is powerful precisely because it acts, and that power cuts both ways.

The fifth limit is price. The full model costs 0.834 dollars per 1M input tokens and 2.501 dollars per 1M output tokens on OpenRouter, which is flagship pricing. A long computer-use session that reads screens, clicks through apps, and finishes multi-step tasks can burn through tokens fast. Hy4 Flash is the budget answer at 0.07 dollars per 1M input tokens, but it is a smaller model with 12B active parameters. If you need the full Tencent Hy4 , budget for it like a flagship, not a free tool.

The sixth limit is language and ecosystem. Hy4 comes from Tencent, a Chinese company, and its deepest integrations are with Tencent apps: CodeBuddy, WorkBuddy, Yuanbao, and WeCom. If your workflow lives in Google Workspace, Microsoft 365, or Adobe tools, the model can still work through the command line and interfaces, but the polished path is the Tencent stack. For a Western office, Tencent Hy4 is a capable guest, not a native.

None of this makes Hy4 a bad model. It makes it a specialised one. The honest picture is a model that trades leaderboard wins for hands-on ability, trades open weights for an API, and trades broad reasoning for deep computer use. If you need a model that finishes tasks inside professional software, those trades are worth making. If you need the strongest general reasoner, the best price, or weights you can download today, Tencent Hy4 is not that model yet.

Tencent Hy4 in the Workplace: Who Should Try It First

Creative professionals should try it first. If your day runs through Blender, CAD tools, Unreal Engine, Unity, or Godot, Hy4 was trained specifically for the software you already use. It reads images and charts, then acts through the command line and interfaces, so a Friday report that used to take an afternoon can start building itself. For this group, Tencent Hy4 is not a curiosity. It is a tool that fits the work.

Office workers should start with Flash. Hy4 Flash is free in WorkBuddy through November 30, 2026, and it keeps the vision and computer use of the full model at a fraction of the price. At 0.07 dollars per 1M input tokens, it costs about one-twelfth of full Hy4, so trying it costs almost nothing. If your work is slides, spreadsheets, and documents, Flash is the sensible first step into the Tencent Hy4 family.

Developers should go through OpenRouter. The model ID is tencent/hy4, or tencent/hy4-flash for the cheaper tier, and the API is the fastest route to try the model in code. Tencent Cloud TokenHub is the alternative, with half price through December 31, 2026. Either way, start with a small task and a sandbox, not a production system. The quickest way to learn what Tencent Hy4 can do is to give it a small job and watch.

Teams with strict data rules should wait. Hy4 through the API means your data leaves your network on every request, and the full model weights are not public, so you cannot run it offline or inspect it. If your work involves client data, health records, or anything regulated, the sandbox rule is not optional. For these teams, Tencent Hy4 is a model to watch, not a model to deploy today.

Students and hobbyists get the most from Flash. The free WorkBuddy offer through November 30, 2026 is a genuine way to try computer use without paying, and the 224B Flash model is still a serious model with a 1M-token context window. Use it to learn how an agent works: give it a multi-step task, watch where it succeeds, and watch where it goes wrong. That lesson is worth more than any benchmark, and Tencent Hy4 Flash makes it free to learn.

Anyone comparing models should try both sides. GPT-6.2 leads on broad reasoning and has the deepest US app ecosystem, while Hy4 leads on hands-on computer work in professional creative software. The two models are not really rivals; they are different tools for different jobs. Try GPT-6.2 for thinking and Tencent Hy4 for doing, and you will see the split clearly. The AI race is no longer about who answers best. It is about who finishes the job.

The practical rule is simple: start small, stay in a sandbox, and scale up only when you trust the model. Give Hy4 a test machine, a copy of your files, and a task you can check, then watch what it does before you give it anything real. That is true for any agent that clicks buttons and runs commands, and it is doubly true for a model as capable as this one. A careful first week with Tencent Hy4 beats a careless first day.

If you run a small business, the office case is the one that matters. A Friday report that builds itself, a slide deck that assembles from a spreadsheet, or a set of charts that land in the right file without you touching the mouse: that is the benefit Tencent is selling. Start with Hy4 Flash in WorkBuddy while it is free, and move to the full model only when the work justifies the price. For a small team, Tencent Hy4 is worth a careful trial, not a blind leap.

What is GPT-6.2 — Astra vs Sol explained, pricing and benchmarks

Why Chinese AI models are removing their hacking skills — Kimi K3.1 Pro, Tencent Hy4 and Qwen4-Max explained

Frequently Asked Questions About Tencent Hy4

What is Tencent Hy4?

Tencent Hy4 is Tencent’s flagship AI model, released November 4, 2026 — a 770-billion-parameter mixture-of-experts model with a 1-million-token context window, and the first Hy model with vision and computer use.

Is Tencent Hy4 free to use?

Partly. Hy4 Flash is free in WorkBuddy through November 30, 2026. Full Hy4 costs $0.834 per 1M input tokens and $2.501 per 1M output tokens on OpenRouter.

Is Tencent Hy4 open source?

Not yet. Tencent plans to release the weights and a system card, but as of mid-November 2026 only the Hy4 preview weights (August 28, Apache 2.0) are public.

What is Hy4 Flash?

Hy4 Flash is the cheaper, faster member of the Hy4 family, released November 16, 2026: 224B total parameters, 12B active per token, a 1M-token context window, and API prices around $0.07 per 1M input tokens — about one-twelfth of full Hy4.

Conclusion: A Model That Works, Not Just Answers

Tencent Hy4 is the first Hy model that does not just think about your work — it does your work. A 770B model that reads your screen and finishes multi-step tasks is a genuine shift, even if the benchmark crown still belongs to Claude and GPT. The smart move: try Hy4 Flash first. At $0.07 per million input tokens, it costs almost nothing to test — and if it saves you an hour a week, the full Hy4 is worth a serious look next.

The takeaway: the AI race is no longer about who answers best. It is about who finishes the job.

Sources: Tencent Hy4 Flash announcement, TechNode on the Hy4 release, ExplainX Hy4 release breakdown.

Leave a Comment