Tencent Hy4 is Tencent’s new flagship artificial intelligence model, released on November 4, 2026, and it is the most important upgrade the company’s Hunyuan model family has ever received. For the first time, a Hunyuan model can genuinely see: it reads images, charts, and documents. And for the first time, it can also use a computer, operating software through a command line or a graphical interface. Tencent says the release puts its model family among the strongest open-weight models in the world for the first time, while openly admitting that a large gap to the best closed models remains. That mix of ambition and honesty is what makes Tencent Hy4 worth a careful look, because the headline and the fine print tell two different stories.
This guide explains what Tencent Hy4 actually does, what it costs in US dollars, where you can try it today, and the one catch every reader should understand before getting excited: despite the open-source framing, the full model’s weights are not downloadable yet. Everything below separates confirmed facts from Tencent’s own claims, so you can judge the release on its merits.
Tencent Hy4 at a Glance
| Detail | What we know |
|---|---|
| Release date | November 4, 2026 (full release; preview arrived August 28, 2026) |
| Maker | Tencent, Hunyuan (Hy) team |
| New in this release | Visual understanding and computer use, both for the first time |
| Reported size | 770 billion parameters, 49 billion active (preview specifications; not restated for the full release) |
| Context window | 1 million tokens (preview specification) |
| API price | $0.834 per million input tokens, $2.501 per million output tokens |
| Discounts live now | 50% off on Tencent Cloud TokenHub through December 31, 2026; free WorkBuddy and CodeBuddy use November 4–15, 2026 |
| Open weights | Not yet released — Tencent says it plans to publish them, with no date given |
| Cheaper version coming | Hy4 Flash, promised soon, with no date or specifications yet |

Contents
- Tencent Hy4 at a Glance
- What Is Tencent Hy4?
- Vision: A Model That Can Finally See
- Computer Use: Clicking and Typing on Your Behalf
- Benchmarks: What Tencent Claims, and What It Does Not Prove
- Tencent Hy4 Pricing in US Dollars
- How to Try Tencent Hy4 Today
- The Open-Weights Catch You Must Understand
- The Cybersecurity Training Decision
- Hy4 Flash and the Two Science Models
- Tencent Hy4 vs the Competition
- Who Should Use Tencent Hy4, and Who Should Wait
- Honest Limitations
- Frequently Asked Questions
- The Verdict
What Is Tencent Hy4?
Tencent Hy4 is the fourth generation of Tencent’s Hunyuan large language model line. Tencent has been rebuilding its AI infrastructure since February 2026, and since then it has shipped a major model roughly every two months. The line grew from fewer than 300 billion parameters in Hy3 to a reported 770 billion in Tencent Hy4, and the company already expects its successor to push further still.
The August 28 preview gave developers an early taste, and it shipped with open weights and a model card on day one. The November 4 full release is a different kind of launch: a bigger, more capable model that is available right now through Tencent’s own apps and through APIs, but whose downloadable weights are still to come. That distinction matters, and this article keeps it visible throughout, because much of the early coverage blurs it.
In practical terms, Tencent Hy4 is a general-purpose model built for long, multi-step work. Tencent highlights three areas of improvement over the preview: instruction following, long-context understanding, and long-horizon agentic tasks. In plain language, the model is supposed to follow complicated instructions more reliably, make sense of very long documents and conversations, and keep working toward a goal across many steps without losing track. Those are exactly the skills that matter when an AI assistant is asked to do real work rather than answer a single question.
One more piece of context helps explain why this launch landed when it did. November 2026 is an unusually crowded month for open-weight models. Moonshot’s Kimi K3.1 Pro opened its weights days earlier, Meituan opened API access to LongCat 2.5, and MiniMax’s M3 flagship arrived on November 3. Tencent Hy4 enters that race with a distinctive pitch: not just a chatbot, but a model that can look at your screen and operate your software.
Vision: A Model That Can Finally See
The single biggest addition in Tencent Hy4 is visual understanding. Earlier Hunyuan models were text-only in practice; this one reads images, charts, and documents as part of its normal work. That sounds simple, but it changes what the model can be used for in a fundamental way.
Consider a financial analyst staring at a quarterly earnings chart, or an engineer reviewing a technical diagram, or a student working through a scanned textbook page. A text-only model needs someone to describe that material first, and every description loses detail. The model can take the image itself, read the values, follow the structure, and reason about what it shows. Tencent has also tuned the model for professional financial analysis, which makes sense: finance is a field where charts, tables, and dense documents are the daily raw material, and where a model that can read them directly has an immediate advantage.
The vision upgrade also feeds the model’s agent skills. An agent that can see a screen, a dashboard, or a document is far more useful than one that can only read text you paste in. It can check its own work against what is actually displayed, notice when a page changed, and handle material that was never written out as plain text. For business users, that is the difference between an assistant that answers questions and an assistant that participates in a workflow.

What vision means for everyday users
For a regular user, the vision capability shows up in Tencent’s own apps first. In WorkBuddy, Tencent’s AI office assistant, and in CodeBuddy, its coding assistant, Tencent Hy4 can now work with visual material alongside text. A user can upload a chart and ask what it means, share a screenshot of an error and ask for a fix, or hand over a document image and ask for a summary. The model still answers in text, but what it can look at has expanded dramatically.
There are sensible limits to keep in mind. Vision models can misread small text, confuse similar-looking values in a dense chart, and describe an image confidently even when they have misinterpreted it. It is no exception to that rule, and anyone using it for financial or technical decisions should verify important numbers against the original source rather than trusting the model’s reading alone.
Computer Use: Clicking and Typing on Your Behalf
The second headline addition is computer use, and it is the more ambitious of the two. The model can operate software through a command line interface or a graphical interface, performing a wide range of work and everyday tasks. In other words, it does not just tell you how to do something; it can take the steps itself, clicking buttons, filling fields, and running commands.
Tencent trained the model specifically on professional software such as Blender and CAD tools, and on game engines including Unreal Engine, Unity, and Godot. That training choice reveals the target audience: not casual users asking a chatbot for recipes, but professionals and developers who work inside complex software all day. A model that understands Blender’s interface or a CAD workflow can help with genuinely skilled tasks, from setting up a scene to navigating a design file.
Computer use is also the capability where caution matters most. An agent that can click and type can also make mistakes at machine speed: deleting the wrong file, sending the wrong message, or changing a setting it misunderstood. The sensible way to use this capability, at least in its first weeks, is on copies of work, in test projects, and with human review before anything irreversible happens. Tencent’s own framing is careful here too: the company presents computer use as a new capability to explore, not as a finished autopilot for your computer.

How computer use differs from a normal chatbot
A normal chatbot lives inside a chat window. You type, it replies, and anything that needs doing in another app is your job. A computer-use model breaks that wall. It perceives the screen, plans a sequence of actions, and carries them out, checking the result of each step before taking the next. That loop of seeing, acting, and checking is what turns a model from an advisor into an agent.
The practical payoff is time. Multi-step chores that eat up an afternoon — gathering data from several tools, reformatting it, and filing it somewhere — are exactly the jobs computer-use models are built to absorb. The practical risk is trust. Every step the model takes on your behalf is a step you did not personally check, so the first deployments should be watched closely, with clear limits on what the agent is allowed to touch.
Benchmarks: What Tencent Claims, and What It Does Not Prove
Tencent’s launch claims are bold, and they deserve to be reported precisely, with their source attached. According to Tencent, the new model outperformed Anthropic’s Claude Opus 5.5 in agentic coding, and surpassed both Claude Opus 5.5 and OpenAI’s GPT 6.1 Sol in agentic search and working-agent tasks. The company also published a table of 34 benchmarks comparing the new model against its own preview, GLM-5.3, DeepSeek V4.1 Flash, and Kimi K3.1 Pro.
Every one of those numbers is company-reported. No independent laboratory has yet re-run the full benchmark suite on the finished model, and independent testing has simply not had time to appear in the first days after launch. That is normal for a model release, but it means readers should treat the claims as Tencent’s best case, not as settled fact. The honest way to read a launch benchmark table is as a statement of intent: this is where the maker believes its model stands, and the wider community will confirm, qualify, or challenge those numbers over the coming weeks.
There is one counterweight to the hype that comes from Tencent itself. The company openly acknowledges a large gap between the new model and the top closed models. That admission is refreshing, and it is also strategically useful: it sets expectations at “among the strongest open models” rather than “the best model in the world.” Readers should hold both thoughts at once. It can be a genuinely strong open model and still sit clearly behind the frontier closed systems, and that is a perfectly respectable place to be.

How to read benchmark claims like a professional
- Ask who ran the test. Vendor-run benchmarks use the vendor’s prompts, settings, and sometimes its own test sets. Independent re-runs are the gold standard, and they take time.
- Ask what the test measures. “Agentic coding” and “agentic search” are broad labels. A model can lead on one benchmark suite and trail on another that measures a slightly different skill.
- Ask about the comparison set. Beating one named rival on one task is not the same as beating every rival on every task. Tencent’s claims are specific, and they should stay specific.
- Wait for the second wave. The most reliable picture of a new model emerges two to four weeks after launch, when independent developers have run it on their own work and reported what actually happened.
Tencent Hy4 Pricing in US Dollars
The new model keeps the same API price as the Hy4 preview. The listed rates are $0.834 per million input tokens and $2.501 per million output tokens, with cached input at $0.042 per million tokens. For readers who think in dollars per task rather than dollars per token: a million input tokens is roughly 750,000 words of text fed to the model, so most individual requests cost a fraction of a cent at these rates, while heavy agent workloads that churn through long documents and many reasoning steps can add up quickly.
Two discounts are live right now, and both have end dates worth noting. From November 4 until December 31, 2026, Tencent Hy4 is billed at 50% of its listed price on Tencent Cloud TokenHub, which brings the effective rate to about $0.417 per million input tokens and $1.251 per million output tokens. Separately, WorkBuddy and CodeBuddy are free to use from November 4 to 15, 2026, a limited free-usage window designed to gather user feedback. If you want to try Tencent Hy4 without paying, that November 4–15 window is the obvious moment.
How does that price compare? The lead research for this article placed Tencent Hy4’s pricing between DeepSeek V4 Pro’s off-peak and peak rates, which puts it in the affordable band for a flagship-class model rather than the premium band. It is not the cheapest model on the market, and it is not trying to be. The pricing logic is straightforward: you pay a mid-range rate for a model that adds vision and computer use, and the launch discounts make the first two months substantially cheaper than the list price.
| Route | Price | Notes |
|---|---|---|
| Tencent Cloud TokenHub (list) | $0.834 input / $2.501 output per million tokens | Cached input $0.042 per million tokens |
| TokenHub launch discount | ~$0.417 input / ~$1.251 output per million tokens | 50% off through December 31, 2026 |
| OpenRouter | ~$0.417 input / ~$1.251 output per million tokens | 50% discount shown at launch; single provider (Tencent Cloud) |
| WorkBuddy / CodeBuddy | Free | November 4–15, 2026 only |
One pricing detail deserves a plain warning. Discounts end. Anyone building a product or a workflow on the model during the 50%-off window should budget for the full list price from January 2027 onward, and should check whether cached-input pricing applies to their usage pattern, since caching is where the biggest savings hide for repetitive agent workloads.
How to Try Tencent Hy4 Today
You can reach the new model today, but the route you choose depends on what you want to do. Here are the confirmed routes, in order of ease for a first try.
- WorkBuddy or CodeBuddy (easiest, free until November 15). Tencent’s own assistants now run on Tencent Hy4. WorkBuddy is the office assistant, CodeBuddy the coding assistant. Open either app, and you are using the new model with no API key, no billing setup, and no code. This is the route to choose if you simply want to see what the model feels like.
- Yuanbao, WeCom’s Dayuan assistant, or ima. The model is live across Tencent’s product ecosystem, so users of these apps can access the model inside tools they may already use.
- Tencent Cloud TokenHub (for developers). The API route for building products. TokenHub carries the 50% launch discount through December 31, 2026, and it is the natural choice for teams already on Tencent Cloud.
- OpenRouter (for developers who prefer one API for many models). The model is listed as tencent/hy4, accepting text and image input, returning text, and supporting tool calling and structured outputs. At launch it had a single provider, Tencent Cloud, so availability depends on that one route.
What you should not do yet is go looking for a download. Despite the open-source framing in much of the coverage, the full model’s weights are not out, and any site claiming to offer a download in the first days after launch should be treated with suspicion. The safe routes are the official apps and the official API listings above.
For readers who already read our earlier coverage, here is the link that ties the story together: our Tencent WorkBuddy guide covered the Hy4 preview and the 9-billion-task office agent in October. This article covers the full November 4 release, which is a different, larger launch, and the two should be read as a before-and-after pair.
The Open-Weights Catch You Must Understand
This is the most important section of the article, because it corrects the single most common misunderstanding about Tencent Hy4. Many headlines describe Tencent Hy4 as an open-source or open-weight model. In practice, right now, it is not yet one you can download and run yourself.
Here is the confirmed position as of November 5, 2026. Tencent says it plans to release the model weights, and it has lined up support from the communities behind vLLM and SGLang, two popular inference engines, plus Inferact. But the company gives no date, and as of the afternoon of November 4, Tencent’s Hugging Face account held only the two Hy4-preview repositories, both under the Apache 2.0 licence. The “Model weights” button on the OpenRouter page linked to a Hugging Face address that was not publicly accessible. The licence for the full release has not been stated at all.
Until those weights ship, every request goes through Tencent’s hosted services. You cannot run the model on your own servers, you cannot fine-tune it on your own data, and you cannot inspect its weights. That is a perfectly normal way to offer a model — most frontier models work exactly this way — but it is not what “open-weight” usually means, and readers who care about self-hosting should wait for the actual release before making plans.

Why the weights delay might not be a problem
For most users, the delay changes nothing. If you use WorkBuddy, CodeBuddy, or an API, you already have the full model today, and the weights question is irrelevant to your daily work. The delay matters mainly to three groups: researchers who want to study the model directly, companies that need to run AI inside their own data centers for privacy or regulatory reasons, and developers who want to fine-tune a private version. If you are in one of those groups, the honest advice is to treat Tencent Hy4 as “coming soon” for your purposes, and to keep an eye on Tencent’s Hugging Face account for the real release.
There is also a constructive way to read the delay. Tencent says the full system card, with fuller safety evaluations, will arrive at the same time as the weights. A system card is the document that explains how a model was tested, where it fails, and what safeguards surround it. Releasing weights and safety documentation together is better practice than dropping weights alone, and it suggests Tencent is taking the safety side of this release seriously — a theme that continues in the next section.
The Cybersecurity Training Decision
One of the most distinctive facts about Tencent Hy4 has nothing to do with benchmarks. According to Reuters, Tencent trained the new model with limited cybersecurity data: material covering exploit development, reverse engineering, and penetration testing was filtered out of the training data, in order to reduce the risk that the model could help with cyberattacks.
This is a deliberate safety trade-off, and it is worth understanding on its own terms. A model trained without exploit-development material will be less useful to someone trying to build an attack, but it may also be less useful to a legitimate security professional trying to defend a system, since defense work often requires understanding exactly how attacks work. Tencent has chosen to err on the side of limiting attack assistance, and that choice tells you something about how the company wants this model to be used: as a general work assistant, not as a security research tool.
The decision also sits inside a wider industry argument. As AI models grow more capable at coding and at operating computers, the question of what they should be allowed to know about hacking becomes more urgent, not less. A model that can use a computer and write code is, in the wrong hands, a more powerful tool than a model that can only chat. Filtering cybersecurity material from training is one answer to that problem. It is not a complete answer — a capable general model can still reason about security topics from general knowledge — but it is a clear, public stance, and it deserves credit as such.
For readers in security roles, the practical consequence is simple: do not choose it as your primary security-testing assistant. Its training was deliberately limited in exactly that area. For everyone else, the decision is mostly reassuring background: the model you are using for office work, coding help, and document analysis was built with a stated safety boundary around attack assistance.
Hy4 Flash and the Two Science Models
Tencent has already announced what comes next. Hy4 Flash, a faster and cheaper variant of the flagship, is promised soon, designed for greater cost-efficiency and high-value everyday applications. No date and no specifications have been given, so treat it as a roadmap item rather than a product you can plan around. The pattern is familiar from the rest of the industry: a flagship model first, then a smaller, faster sibling for the workloads where speed and price matter more than maximum capability.
Less widely reported, but genuinely interesting, is that Tencent’s Hy team also open-sourced two science models built with the flagship model and its Hyra 2.0 harness: HyCell, for single-cell representations, and HyPDE, for physics fields. These are specialist research tools, not consumer products, but they show the model family being put to work in real scientific domains rather than only in chatbots and office assistants. For researchers in biology or physics, they may turn out to be the most useful part of the whole release.
Tencent Hy4 vs the Competition
Placing the new model beside its rivals helps clarify what kind of model it is. The table below uses only claims and prices that are sourced in this article, and it keeps vendor claims labelled as vendor claims.
| Model | What stands out | Price signal |
|---|---|---|
| Tencent Hy4 | Vision + computer use for the first time; strong agentic claims (vendor-reported); weights promised, not yet out | $0.834 / $2.501 per million input/output tokens; 50% off into year-end |
| Claude Opus 5.5 (Anthropic) | The named rival in Tencent’s agentic-coding and agentic-search claims; closed model | Premium closed-model pricing (see Anthropic’s current rate card before comparing) |
| GPT 6.1 Sol (OpenAI) | The second named rival in Tencent’s agentic-search and working-agent claims; closed model | Premium closed-model pricing |
| DeepSeek V4.1 Flash | Efficiency-first rival; Tencent’s own 34-benchmark table includes it as a comparison point | Lower-cost band; Tencent Hy4 sits between its off-peak and peak rates per lead research |
| Kimi K3.1 Pro (Moonshot) | Weights opened days before Tencent Hy4; direct competitor on coding and agent work | Open weights available now, unlike Tencent Hy4’s full release |
| LongCat 2.5 (Meituan) | API access opened days before Tencent Hy4; plugs into Claude Code workflows | Open-weight challenger in the same November window |
The comparison reveals the real contest of November 2026. It is not one model against the world; it is a cluster of open-weight and open-weight-promised models, each pitching a different strength. Kimi K3.1 Pro wins on weights actually being available. LongCat 2.5 wins on developer-workflow integration. The new model wins, on paper, on the combination of vision and computer use in a single model. Which of those strengths matters most depends entirely on what you want to build.
Readers who want the wider November picture can follow the cluster through our recent coverage: the Drex Max open-weights guide for a decision-model angle, the LongCat 2.5 explainer for Meituan’s challenger, and the DeepSeek Super Agent guide for the agent-workflow story that Tencent Hy4 now joins.
Who Should Use Tencent Hy4, and Who Should Wait
Try it this week if you are…
- A WorkBuddy or CodeBuddy user. You already have access, and the November 4–15 free window costs nothing. Try the vision features on real charts and documents from your own work.
- A developer building agent workflows. The OpenRouter and TokenHub routes give you tool calling and structured outputs today, at a 50% launch discount, which is a cheap way to test whether the agentic claims hold up on your tasks.
- A finance or data professional. The financial-analysis tuning and the vision upgrade make this the most relevant Tencent model yet for chart-heavy, document-heavy work — with the verification caveat below.
- An AI watcher tracking the open-model race. Tencent Hy4 is a major data point in the November story, and using it first-hand beats reading about it second-hand.
Wait if you are…
- Planning to self-host or fine-tune. The weights are not out, the licence is unstated, and no date has been given. Wait for the real release.
- Building a product that depends on a fixed price. The 50% discount ends December 31, 2026, and the free app window ends November 15. Budget at list price.
- Doing security research or penetration testing. Cybersecurity training data was deliberately limited. Choose a model built for that work instead.
- Needing guaranteed accuracy on critical numbers. Vision readings of dense charts should always be verified against the source, especially in the first weeks.
Honest Limitations
No launch article is complete without the other side of the story, so here it is, plainly stated. First, the benchmark claims are Tencent’s own, and independent verification has not had time to appear. Second, the weights are promised, not delivered, and the licence is unstated — so the “open” label is aspirational for the full release today. Third, Tencent itself acknowledges a large gap to the best closed models, which means buyers should not expect frontier-leading performance on every task. Fourth, computer use is a new capability, and new agent capabilities fail in new ways: misunderstood instructions, wrong clicks, and confident errors are all part of the early experience. Fifth, the cybersecurity training filter, while defensible, limits the model’s usefulness for legitimate security work.
None of those limitations makes Tencent Hy4 a bad model. They make it a new model, launched quickly into a crowded month, with some genuinely new capabilities and some genuinely unfinished business. The right response is curiosity with verification: try it on real work, check its answers, and judge it by what it does for you rather than by its launch table.
Frequently Asked Questions
Is Tencent Hy4 really open source?
Not yet in practice. Tencent says it plans to release the weights and a system card, but no date has been given, and the licence for the full release is unstated. The August preview was open-weight under Apache 2.0; the November full release is currently available only through Tencent’s apps and APIs.
How much does Tencent Hy4 cost?
The list price is $0.834 per million input tokens and $2.501 per million output tokens, with cached input at $0.042 per million tokens. A 50% discount applies on Tencent Cloud TokenHub through December 31, 2026, and WorkBuddy and CodeBuddy are free from November 4 to 15, 2026.
What is new in Tencent Hy4 compared with the preview?
Two things: visual understanding and computer use, both for the first time in the Hunyuan line, plus improvements in instruction following, long-context comprehension, and long-horizon agent tasks. The preview, released August 28, 2026, was text-focused by comparison.
Can Tencent Hy4 use my computer?
It can operate software through a command line or a graphical interface, and it was trained on professional tools such as Blender, CAD software, Unreal Engine, Unity, and Godot. Use it on test projects first, review its actions, and keep it away from irreversible operations until you trust it.
Is Tencent Hy4 better than Claude Opus 5.5 or GPT 6.1 Sol?
Tencent claims it outperformed Claude Opus 5.5 in agentic coding and surpassed both rivals in agentic search and working-agent tasks, but those are company-reported results, not independently verified ones. Treat them as a claim to be tested, not a verdict.
When is Hy4 Flash coming?
Tencent says soon, with no date or specifications. It is designed to be cheaper and faster than the full Tencent Hy4, aimed at high-value everyday applications.
Should I wait for the weights before trying Tencent Hy4?
Only if you need to self-host, fine-tune, or run the model inside your own data center. For everyone else, the apps and APIs already give you the full model today, and the November free window is the cheapest time to try it.
The Verdict
The new release is a serious one with an asterisk, and both halves of that sentence matter. The serious part is real: vision and computer use in one model, strong agentic claims, a mid-range price with generous launch discounts, and a free trial window that runs until November 15. For office workers, developers, and data professionals, there is something genuinely new to try this week, and the cheapest way to try it costs nothing at all.
The asterisk is real too. The weights are not out, the licence is unstated, the benchmarks are vendor-reported, and Tencent itself concedes the gap to the closed frontier. None of that should stop you from trying the new model. All of it should stop you from building long-term plans on it before the weights, the system card, and the independent tests arrive.
If you take one action from this article, make it this: open WorkBuddy or CodeBuddy before November 15, hand the new model a real chart or a real multi-step task from your own work, and see what it does. The November open-model race will be decided by that kind of hands-on evidence, not by launch tables — and the model, for all its unfinished business, has earned its place in the test.
Sources for this article include Tencent’s November 4, 2026 launch announcement as reported by PYMNTS and by ShiftDelete, Reuters’ reporting on the model’s cybersecurity training data, and live API listings on Tencent Cloud TokenHub and OpenRouter, all checked on November 5, 2026. Benchmark figures are Tencent’s own and are labelled as such throughout.
