Tencent Hy4 Is 770 Billion Parameters That Can See and Use a Computer: $0.834 Per Million Tokens, Free Until 15 November

What is Hy4, and what is actually new?

Hy4 is Tencent's new flagship model in the Hunyuan line. The headline number is scale: 770 billion total parameters in a mixture of experts design, which means the model routes each token to a subset of experts rather than waking the whole model every time. That is a 2.6 times increase over Hy3's 295 billion parameters, and it follows the Hy4 preview from late August 2026, which already carried the 770 billion total. Scale alone is not the story. The story is the two capabilities that are new to the Hunyuan line. First, vision. Hy4 can interpret charts and documents, reason across multiple images, and handle visual mathematical reasoning, which means it can look at a chart, a screenshot, or a document image and answer from what it sees, not just from text you type. Second, computer use. Hy4 can operate a computer through its graphical interface and command line tools, combining perception, reasoning, and tool calls to carry out tasks across programs. Tencent says targeted training covers professional software such as Blender and CAD tools, plus office tasks, financial analysis, and game development, with product teams helping train and evaluate the model against real workflows. That combination is what makes Hy4 interesting for readers who build things, not just for readers who follow model releases. A model that can read a chart and then act in an app can, in principle, take a screenshot of a broken layout, understand what is wrong, and then change the file, instead of stopping at a description. In practice, computer use models still need supervision, limits, and testing, and Tencent's own admission about long horizon coding shows this is not a finished, hands off agent. The useful question is not whether Hy4 can use a computer in a demo, but whether it can finish your specific multi step task without breaking something else, and that only a test on your own task can answer.

What does Hy4 cost, with and without the discount?

The pricing story is unusually clean, because Tencent kept the list price unchanged from the preview. The standard API price is $0.834 per million input tokens and $2.501 per million output tokens. In Chinese pricing reported from Tencent's release, that is 6 yuan per million input tokens, 18 yuan per million output tokens, and 0.3 yuan per million cache hit tokens. This article quotes the US dollar list price as the main figure, because that is the price most readers outside China will compare, and it does not convert the yuan cache price into dollars, because no official conversion is published in the reporting checked. The discount is on Tencent Cloud TokenHub only, and it is time limited. From launch until 31 December 2026, Hy4 is charged at a 50 percent discount on TokenHub, based on the published price. Halving the list price gives $0.417 per million input tokens and $1.2505 per million output tokens on TokenHub during the discount window. Those discounted figures are calculations from the stated list price and the stated 50 percent discount, not separate prices quoted by Tencent, and they apply on TokenHub, not automatically on every gateway. | Route and period | Input per million tokens | Output per million tokens | What to check | | Standard API list price, from 4 November | $0.834 | $2.501 | Unchanged from the Hy4 preview | | Tencent Cloud TokenHub, until 31 December 2026 | $0.417, calculated as half the list price | $1.2505, calculated as half the list price | Discount is TokenHub only, and ends 31 December | | Third party gateways such as OpenRouter | Varies by gateway | Varies by gateway | Gateway may add its own margin, routing, or fees, so check the gateway price, not just the Tencent list price | | WorkBuddy and CodeBuddy, 4 to 15 November | Limited free quota, amount not stated | Limited free quota, amount not stated | Free amount is not published in the reporting checked, so check inside the app before planning a job | A worked example shows why the discount matters less than the task design. Suppose a document task uses 10 million input tokens and 2 million output tokens. At list price, that is 10 times $0.834, plus 2 times $2.501, which is $8.34 plus $5.002, or $13.342 in total. At the TokenHub discount, the same task is $4.17 plus $2.501, or $6.671 in total. The discount saves about $6.67 on that job, which is real, but it does not turn a bad task design into a good one. A task that feeds the same large document on every step will still cost more than a task that caches, summarises, or targets the input, at either price.

Where can you actually use Hy4 today?

Hy4 is available now as a hosted service, not as a download you can run yourself. Tencent lists it across its own products: Tencent Cloud TokenHub, WorkBuddy and CodeBuddy, Yuanbao, ima, and the WeCom intelligent assistant Dayuan. It is also available through third party API platforms, including OpenRouter, Hermes, and OpenCode, in the reporting checked. That split matters, because the product route and the API route are different ways to use the same model. The product route is the fastest test. If you already use Yuanbao, WorkBuddy, or CodeBuddy, Hy4 should appear as a model choice or as the engine behind a task, and the limited free quota from 4 to 15 November is the window to test it without paying. Use that window for a real task you already have, such as reading a set of charts, summarising a document with images, or automating a small multi app workflow, and keep a note of where it succeeded, where it needed correction, and where it stopped. That note is worth more than any benchmark table, because it is about your work. The API route is for developers. OpenRouter lists Hy4 with a 1 million token context window in its gateway listing, and other gateways list it as an OpenAI compatible endpoint. Treat the 1 million context figure as a gateway listing figure, not as a vendor confirmed spec from Tencent's own announcement material, because the coverage checked flags it that way. Before building a product on a 1 million token claim, test the actual context behaviour through your chosen gateway, including how it handles long documents, many images, and long computer use sessions, because gateway limits and model limits are not always the same. The route that is not available yet is self hosting. Tencent says it plans to release the weights shortly, with support from vLLM and SGLang and a system card with fuller safety results, but it gave no date in the reporting checked. As of launch morning, only the August preview was on Tencent's Hugging Face page, under the Apache 2.0 licence for the preview, and Tencent has not said which licence the full Hy4 weights will use. A cheaper, faster Hy4 Flash is also due soon, with no specs or date published. Until the weights and Flash actually ship, any plan that depends on running Hy4 yourself, or on a Flash price, is a plan based on a promise, not a product. Build on the hosted API or the Tencent apps today, and treat weights and Flash as future options to check again.

What does Tencent claim, and what should you not take on trust?

Tencent's framing, as reported by Nikkei Asia and others, is that Hy4 breaks into the top tier of open source models for the first time. ChainCatcher's report of Tencent's release says Hy4 leads in evaluation across scenarios such as finance, office, development, and scientific research, and that it has been tested on four major practical tasks. Other coverage says Hy4 outperformed its late August preview on 12 tasks covering agent capabilities, programming, and work automation. Those are Tencent's own evaluation results, from its own technical blog and release material, and they should be read that way until independent reviewers test the full model on the same tasks. The more useful honesty in the launch is what Tencent admits Hy4 cannot yet do well. Reporting of Tencent's own material says Hy4 lags behind top closed source systems in research level mathematics, long horizon coding, model self improvement, and advanced visual question answering. That admission sets the right expectation for readers. Hy4 is a strong, cheap, vision and computer use model for office, finance, development, and game workflows that Tencent trained it for, not a proven replacement for the best closed models on the hardest research or coding tasks. If your task is in the admitted weak areas, test Hy4 against a model you already trust on the same task, side by side, before switching. There is also a pricing trap to avoid. The 50 percent TokenHub discount ends on 31 December 2026, and the WorkBuddy and CodeBuddy free quota ends on 15 November 2026. A service, a workflow, or a client quote that assumes the discount price forever, or assumes free use after 15 November, will break in January. Price your work at the list price, treat the discount as a margin bonus during the window, and check the live TokenHub price before you quote a job that runs past 31 December.

How should you test Hy4 this week?

Start with a vision test that has a right answer you can check. Give Hy4 a chart with a clear trend, a document image with a table, and a screenshot with an error message, and check whether its reading matches the numbers you can see yourself. Vision models fail in quiet ways, such as misreading a chart axis or a table row, so a test with a known answer catches those failures before they reach a client. Then test a computer use task with a safe boundary. Pick a task in a test folder or a test account, with no live data, no real payments, and no irreversible deletes, and let Hy4 work through the graphical interface and command line steps while you watch. Note every step where you had to correct it, and every step where it stopped and asked for help. A computer use agent that needs correction on step 2 of a 10 step task is not ready for unsupervised use, even if it finishes the demo. Then test the cost on your real token mix. Run one real job through TokenHub during the discount, record input tokens, output tokens, and the bill, and compare it with the same job on a model you already use. One real bill tells you more than a pricing table, because your job's input to output ratio decides whether Hy4 is cheap for you. A vision heavy job with large image inputs will price differently from a text heavy job with small outputs, even at the same per million rates. For readers comparing large models, our guide to Mistral Large 4 at $0.70 per million tokens and our guide to GPT-6.1 Terra computer use at $1 per million tokens put Hy4's price and computer use in context, and our guide to Kimi K3.1 Pro at $2 per million tokens shows another agent priced model to compare it with. Sources: Tencent Hy4 launch, 4 November 2026, as reported by TechNode, 4 November 2026, for launch date, vision and computer use, 770 billion parameters, $0.834 and $2.501 list prices, 50 percent TokenHub discount until 31 December, and WorkBuddy and CodeBuddy free quota 4 to 15 November; ChainCatcher, 4 November 2026, for 6, 18 and 0.3 yuan pricing and TokenHub discount wording; Nikkei Asia, 4 November 2026, for Tencent's top tier claim; independent launch analysis, 4 November 2026, for weights not yet released, Hy4 Flash with no date, 1 million token context as a gateway listing, and Tencent's admitted gaps in research maths, long horizon coding, self improvement, and advanced visual question answering.