Tencent Hy4 Explained: 770B AI That Sees and Uses a Computer
Tencent just gave its flagship AI model eyes and hands. On November 4, 2026, the company's Hunyuan team released the full version of Tencent Hy4, a 770-billion-parameter AI model that can read images, reason over charts, and operate a computer's command line and graphical interface on its own. It is the most interesting open-model launch of the week, and also one of the easiest to misunderstand — because the headline feature everyone wants, downloadable model weights, is not actually out yet.
Hy4 matters for a simple reason: it moves the open-model race from "can it write good text?" to "can it do real work on a real computer?" Tencent trained it specifically for office tasks, financial analysis, and game development, including professional software like Blender and CAD tools. That focus, plus a one-million-token context window, makes it a productivity model first and a chatbot second. Here is what it actually does, what the benchmarks really say, and how you can try it today.
Contents: What Is Tencent Hy4? · What's New vs the Preview · How the Vision Works · How Computer Use Works · Benchmarks, Honestly Read · Hy4 vs Rivals · How to Try Hy4 · Who It's For · Pros and Cons · FAQs · Verdict
What Is Tencent Hy4?
Tencent Hy4 is the fourth generation of Tencent's Hunyuan model family and the company's new flagship for agentic, productivity-focused work. The "Hy" in the name is short for Hunyuan, Tencent's AI brand, and Hy4 is the successor to the Hy3 line, which topped out below 300 billion parameters. Hy4 jumps to 770 billion total parameters — but, crucially, it does not use all of them at once.
Hy4 is a mixture-of-experts (MoE) model. Think of it as a company with 256 specialist teams: for every word it processes, a router picks just the eight most relevant experts, plus one shared expert, and only those run. That is how a 770-billion-parameter model ends up activating only about 49 billion parameters per token — big knowledge, manageable compute. The same design idea powers several of the strongest open models right now, including DeepSeek V4.1 Flash and MiniMax M3.1 Flash, and it is a big reason open models keep closing the gap on closed ones.
One honesty note up front: the 770B total, 49B active, and one-million-token context figures come from the August preview's model card, and several reports note that Tencent's November announcement did not restate them. They are the best figures on the record, and they are consistent across the preview's documentation, but until the full model's own system card lands, treat them as carried-over preview specs rather than freshly confirmed final specs.
What's New in the Full Hy4 vs the August Preview?
The August 28 preview was text-only. It was built for coding and office work through tool schemas — structured instructions, not a live desktop. The November 4 full release adds two firsts for the Hunyuan line, plus much heavier training behind them:
- Visual understanding. Hy4 can now take images as input, read fine details, analyze complex charts, and solve math problems presented visually.
- Computer use. Hy4 can operate both command-line tools and graphical interfaces, chaining steps across applications to finish long tasks.
- More training, same price. Tencent says it substantially scaled up pre-training and post-training versus the preview, while keeping API pricing unchanged.
- Better instruction following and long-task performance. The release emphasizes long-horizon agent work — tasks that span several apps and many steps, not single replies.
In plain terms, the preview was a very strong text brain. The full Hy4 is that brain with eyes and hands attached.
How the Vision Works — and Why the Demo Matters
Vision in Hy4 is not a bolted-on image captioner. Tencent describes visual reasoning built into the core model, so it can combine what it sees with tool calls and multi-step reasoning. The demo Tencent chose is telling: the model examined a 2.8-gigapixel aerial photograph and located a roof component just two metres across. That is not a party trick — it is the exact skill that matters for reading dense documents, spotting a wrong number in a financial chart, or checking a design render for a misplaced element.
For everyday users, the practical uses are more ordinary and more useful: photograph a whiteboard and ask for the action items, drop in a chart and ask what changed between quarters, or screenshot an error message and ask what to do next. Creators can use the same skill in reverse — studying what makes a thumbnail work — and if you want to grab a competitor's thumbnail to analyze it, our free YouTube Thumbnail Downloader does that in seconds. Vision models are only as good as the images you feed them, so keep files reasonably sized first with our Image Compressor before uploading huge originals.
How Computer Use Works
Computer use is the feature that separates Hy4 from most open models. Instead of only answering in text, Hy4 can look at a screen, decide what to click or type, run commands, and check whether the result looks right before moving on. Tencent gave it dedicated training on professional software such as Blender and CAD tools, and on long tasks that cross several applications — the kind of work where a human would juggle a spreadsheet, a document, and a design tool in one sitting.
That is also where expectations need managing. Computer-use agents are impressive in demos and uneven in real life: a changed button label, an unexpected pop-up, or a permission dialog can stall them. Tencent's own benchmark story reflects that reality — Hy4 is close to the best closed models on computer-use tests, not ahead of them. Treat it as a junior operator that needs supervision, not an autopilot you can walk away from, especially for anything involving money, files you cannot replace, or accounts you care about.
Benchmarks: What Hy4's Numbers Actually Say
Every number in this section comes from Tencent's own launch materials, as reported by independent coverage. None of it has been independently reproduced yet, because the weights are not public. Read them as "Tencent's claims, awaiting verification," not as settled fact.
| Benchmark | Hy4 | Best rival in Tencent's table |
|---|---|---|
| OSWorld 2.1 (computer use) | 78.9 | Claude Opus 5.5 — 81.8 |
| Finance blind evaluation (579 real tasks, 213 professionals) | 3.04 | DeepSeek-V4.1-Flash — 2.97 |
| OSWorld 2.1, other open models | 78.9 | Kimi K3.1 Pro — 79.6; GPT-6.1 Sol — 75.2 |
Two readings are fair at once. First, Hy4 genuinely sits in the top tier of open models: Tencent says the release puts the Hunyuan family among the strongest open models for the first time, and the computer-use score backs that up — it trails Claude Opus 5.5 by less than three points and sits essentially level with Kimi K3.1 Pro. Second, Tencent itself acknowledges a "large gap" between Hy4 and the top closed models overall, and its 34-benchmark appendix mixes wins and losses rather than a clean sweep. The finance result is the most interesting single claim, because it was a blind evaluation by professionals on real tasks rather than a synthetic test — but again, it is Tencent's evaluation, and the scoring scale is Tencent's own.
The broader pattern is worth noting: the open-model field is now a genuine multi-way race. Kimi, DeepSeek, MiniMax, and now Tencent are all shipping models that, on the tests their makers choose, land within a few points of the closed frontier. That competition is why API prices keep falling and why the "open vs closed" gap is now measured in tasks, not in leagues.
Tencent Hy4 vs the Rivals
| Model | Total parameters | Vision + computer use | Weights status (Nov 5) | API input price |
|---|---|---|---|---|
| Tencent Hy4 | 770B (49B active) | Yes, both new | Promised, not yet released | $0.834 / 1M tokens |
| Kimi K3.1 Pro | 2.8T (104B active) | Coding agents; computer use via Kimi Code | Promised by Nov 4 | $2 / 1M tokens |
| DeepSeek V4.1 Flash | Open-weight class | Text-focused flagship | Open weights available | Varies by provider |
| Claude Opus 5.5 | Not disclosed | Yes, leading computer use | Closed | Closed pricing |
Hy4's sharpest edge is price. At $0.834 per million input tokens and $2.501 per million output tokens — and a 50% launch discount on Tencent Cloud TokenHub through December 31, 2026 — it undercuts most rivals for heavy document and agent workloads, where output tokens and long contexts dominate the bill. That is the same pricing logic that made OpenAI's GPT-6 with its intelligent UI and Google's Gemini 4 Argon interesting to compare on value rather than raw scores: the best model is often the one you can afford to run all day.
How to Try Tencent Hy4 Today
Because the full model's weights are not downloadable yet, the only route in is through an API or a Tencent product. Here is the practical path:
- Pick your route. Developers should start with Tencent Cloud TokenHub or OpenRouter, where Hy4 is listed as an API model. Non-developers will find it easier inside Tencent's WorkBuddy or CodeBuddy products, or in Yuanbao.
- Use the free window. WorkBuddy and CodeBuddy are offering a free usage quota from November 4 through November 15. Try your hardest real task there first, before spending anything.
- Test vision with a real document. Upload a chart, invoice, or multi-page report and ask specific questions with checkable answers — "What was revenue in Q3?" beats "Summarize this."
- Test computer use in a sandbox. Give it a disposable environment and a reversible task first. Do not hand it your main machine, your email, or payment accounts until you have seen how it handles surprises.
- Watch the price. After the free quota ends on November 15, standard pricing is 6 yuan per million input tokens and 18 yuan per million output tokens (about $0.834 and $2.501). TokenHub's 50% discount runs through December 31, 2026.
- Wait for the weights before self-hosting plans. Tencent says the weights and system card will follow "in the coming days" with no date. Until they appear, every request goes through Tencent's API, so Hy4 cannot yet run as local AI on your own servers.
Who Tencent Hy4 Is Actually For
Finance and office teams get the most direct value: long documents, spreadsheets, and charts are Hy4's home turf, and the one-million-token context means an entire quarter's reporting can sit in one conversation. Developers and game studios get a coding and agent model trained on Blender and CAD workflows, which is rarer than generic coding ability. Researchers get a model built for long, multi-step work across documents and tools. And AI builders on a budget get frontier-adjacent performance at a price that makes heavy agent loops affordable.
It is a weaker fit if you need a fully offline, private deployment today — that door only opens when the weights ship — or if you want a polished consumer chatbot. Hy4 is a workhorse API model, not a lifestyle app, and Tencent's own products around it are still catching up to the model itself.
Pros and Cons
Pros: vision and computer use are genuinely new for this model family; one-million-token context; strong computer-use and finance results on Tencent's tests; very competitive API pricing with a launch discount; free quota through November 15; open weights promised.
Cons: weights, licence, and system card not yet published, so "open" is a promise, not a fact, for the full model; all benchmarks are company-reported and unverified; Tencent concedes a gap to the top closed models; computer-use agents still need supervision; core specs are carried over from the preview, not restated for the final model.
Frequently Asked Questions
Is Tencent Hy4 free to use?
Partly. WorkBuddy and CodeBuddy offer a free usage quota from November 4 to November 15, 2026. After that, API use costs about $0.834 per million input tokens and $2.501 per million output tokens, with a 50% TokenHub discount through December 31.
Is Tencent Hy4 really open source?
Not yet for the full model. The August preview shipped under Apache 2.0, but the November full release is API-only so far. Tencent says it plans to release the weights and system card in the coming days, with no date and no licence confirmed for the final version.
What is Tencent Hy4's computer use feature?
Computer use lets Hy4 operate command-line tools and graphical interfaces itself, clicking, typing, and checking results across applications. Tencent trained it on professional software like Blender and CAD tools for long, multi-step tasks.
How does Tencent Hy4 compare to Claude Opus 5.5?
On Tencent's OSWorld 2.1 computer-use test, Hy4 scores 78.9 against 81.8 for Claude Opus 5.5 — close, but behind. Tencent also acknowledges a large gap to the top closed models overall, so Hy4 leads on price, not on every benchmark.
When will the Tencent Hy4 weights be released?
Tencent has only said "in the coming days" with no specific date. Until the weights appear, Hy4 can only be used through Tencent Cloud TokenHub, OpenRouter, or Tencent products like WorkBuddy, CodeBuddy, and Yuanbao.
Can Tencent Hy4 run on my own computer?
No, not yet. Running a 770B-parameter model locally needs serious multi-GPU hardware anyway, and the full model's weights are not downloadable. For now, use the API, and revisit self-hosting once the weights and deployment guides ship.
Verdict: A Serious Open-Model Step, With One Asterisk
Tencent Hy4 is the rare launch that is both genuinely exciting and genuinely unfinished. The vision and computer-use additions are real, the pricing is aggressive, and the computer-use score puts it within touching distance of the best closed model on that one test. But the story's most important word is "promised": until the weights, licence, and system card actually ship, Hy4 is an open model in intent and a closed API in practice.
That is not a reason to ignore it — it is a reason to try it the smart way. Use the free quota before November 15, test it on your hardest real document or workflow, and judge it on what it does for you rather than on Tencent's benchmark table. If the weights land as promised, Hy4 could become the default budget agent model for a lot of teams. If they slip, the API alone is still cheap enough to matter. Either way, the open-model race just got a lot more interesting, and Tencent is now unmistakably in it.
For more on where open models are heading, read our coverage of Meta's open Muse Spark model and Anthropic's Claude Duet agent — and if agents doing your computer work sounds useful, Notion's rebuilt AI agents show the same idea inside everyday productivity software. You can also explore Google's Nano Banana Pro image generator for the vision side of the same AI wave.
Last updated: November 5, 2026. Research methodology: this article is based on Tencent's November 4, 2026 launch materials as reported by Startup Fortune, Maccus, ChainCatcher, and UnicornAPI, plus the August Hy4 preview model card for carried-over specifications. Benchmark figures are company-reported and have not been independently verified. No sponsorship or payment influenced this article. Learn more about us or contact us. Browse more AI News on The AI Tools.
Comments
Post a Comment