Tencent Hy4 Is Live: 770B Model That Sees and Uses Your PC

Tencent's flagship AI model can finally see your screen and click around on it. Hy4, released on November 4, 2026, adds visual understanding and computer use to the Hunyuan line for the first time, and Tencent says it now ranks among the strongest open-weight models in the world. There is one large catch: the model itself is not actually open yet. The weights Tencent promised "in the coming days" have still not appeared, and as of November 8 the only thing you can download is the August preview. What you can do is call the hosted model through Tencent Cloud or OpenRouter, where it currently costs half its list price. This review covers what Hy4 does, what Tencent's own benchmarks show, what it costs, and where it still falls well short of the closed models it is chasing.

What Hy4 Is

Hy4 is the latest flagship in Tencent's Hunyuan (Hy) model family, the successor to the Hy4 preview that shipped in late August 2026 as a text-only model. It is a mixture-of-experts design: the preview's model card lists 770 billion total parameters with 49 billion active per token, and launch coverage describes the full release on the same architecture, about 2.6 times the 295 billion parameters of Hy3. Tencent's own release page does not restate the parameter count for the full model, so treat the 770 billion figure as the preview card's specification rather than a number Tencent re-confirmed on November 4.

The context window is listed at 1,048,576 tokens on OpenRouter, with up to 65,536 output tokens. The model accepts text and images as input and returns text, and it supports tool calling and structured outputs. It is live now as a hosted service on Tencent Cloud TokenHub, WorkBuddy, CodeBuddy, the Yuanbao chatbot, WeCom's AI assistant Dayuan, and the ima app, plus third-party access through OpenRouter, Hermes, and OpenCode. It is not currently available through OpenRouter's EU in-region routing.

Two things are new in the full release. First, visual understanding: Hy4 can read images, interpret charts and documents, and reason across multiple images, including visual maths. Second, computer use: it can operate command-line and graphical interfaces, and Tencent says it received specialized training on professional software such as Blender and CAD tools, plus game engines including Unreal Engine, Unity, and Godot.

What Tencent's Own Benchmarks Show

Every score below comes from Tencent's own release page, and several come from tests Tencent built itself. No independent scoreboard has published a full Intelligence Index entry for Hy4 yet. The honest way to read these numbers is as Tencent's best case, not as a neutral verdict.

On computer use, the headline test is OSWorld 2.1, where Hy4 scores 78.9 against Claude Opus 5.5 at 81.8, Kimi K3.1 Pro at 79.8, GPT-6.1 Sol at 75.2, and DeepSeek-V4.1-Flash-1017 at 66.4. That is close to the frontier, and it is the strongest part of the release. On coding, the picture is weaker. Terminal Bench 4.0 puts Hy4 at 44.4, behind Kimi K3.1 Pro at 66.8, Claude Opus 5.5 at 66.4, and GPT-6.1 Sol at 55.5, though ahead of GLM-5.3 at 42.9 and DeepSeek at 40.0. ProgramBench is worse still: Hy4 at 30.0 against Claude at 56.5 and GPT-6.1 Sol at 46.0. Agents' Last Exam v1.1 has Hy4 at 41.0, just behind Claude at 43.8 and Kimi at 43.4.

Tencent also ran a blind finance evaluation with 213 financial professionals across 579 real-world tasks, in partnership with Tencent Financial Technology. Hy4 scored 3.04, ahead of DeepSeek-V4.1-Flash-1017 at 2.97 and Kimi K3.1 Pro and GLM-5.3 at 2.68 each. That is a Tencent-run test on Tencent's home turf, so read it as a signal about finance workflows rather than a general ranking. Our Kimi K3.1 Pro review and DeepSeek V4.1 Flash 1017 review cover the two rivals in more detail.

The one independent aggregation available so far is BenchLeader, which scores Hy4 at 55.8 on its index as of November 8, strongest in maths (66) and weakest in agents and tools (52), with an Artificial Analysis HLE-Diamond result of 32.7%. That is a useful reality check against Tencent's own chart: solid, mid-pack, and a long way from the top of the board.

One more caution on the numbers. Tencent's own page is not fully self-consistent: its appendix table lists slightly different scores from its overview chart for a few benchmarks, including Terminal Bench Science. Treat every figure in this section as Tencent-reported and subject to revision, not as independently verified.

Where It Still Trails

Tencent deserves credit for publishing a limitations chart alongside the wins, because the gaps are large. On its internal Hy-FrontierMath test, Hy4 scores 13.0 against Claude Opus 5.5 at 91.3 and GPT-6.1 Sol at 97.5. On Terminal Bench Science, Hy4 scores 16.9 against GPT-6.1 Sol at 80.9 and Claude at 66.2. On Goal2Repo, a test of turning a goal into a working repository, Hy4 scores 23.2 against Claude at 74.0 and GPT-6.1 Sol at 58.0. Tencent says plainly in its announcement that Hy4 still lags the top closed models on research-level mathematics, long-horizon coding, model self-improvement, and advanced visual question answering.

The pattern is consistent: Hy4 is competitive at operating a computer and at everyday office and finance work, and clearly behind on the hardest reasoning and science tasks. That matters for how you use it. Driving a GUI, filling in a spreadsheet, or building a standard web page are the jobs it is built for. A multi-day research project or a novel proof is not.

What Hy4 Costs

Hy4 keeps the preview's prices, which is the most user-friendly part of the launch. The list price is 6 yuan per million input tokens, 18 yuan per million output tokens, and 0.3 yuan per million cached input tokens. Converted at Tencent's own published dollar figures, that is $0.834 per million input tokens, $2.501 per million output tokens, and $0.042 per million cached tokens. Tencent's press release rounds those figures up to approximately $0.90, $2.70, and $0.045, so if you see both sets of numbers, the smaller ones are the precise conversion and the larger ones are Tencent's own rounding.

Two discounts are running right now. On Tencent Cloud TokenHub, Hy4 is 50% off its standard pricing through December 31, 2026. On OpenRouter, the model is listed at exactly half the list price: $0.417 per million input tokens, $1.251 per million output tokens, and $0.021 per million cached tokens, with a single provider, Tencent Cloud. OpenRouter reports 48 tokens per second throughput, 3.27 seconds latency, and 100% uptime over the past three days. Treat the OpenRouter half price as a possible promotion rather than a permanent rate; the preview model on the same platform still lists at the full $0.834 and $2.501.

Converted at November 8, 2026 mid-market rates of 1 USD = 0.7471 GBP and 1 USD = 1.41 CAD; sterling and Canadian figures are approximations. There is also a free route for the next week: WorkBuddy and CodeBuddy are free to use from November 4 to 15, 2026. A cheaper variant, Hy4 Flash, is scheduled for release soon, according to Tencent, but no price or date has been announced.

For context, Hy4's list price sits between the budget and premium tiers of its rivals. DeepSeek V4.1 Flash 1017 costs $0.15 per million input tokens, so Hy4 is more than five times dearer on input, while Gemini 4 Argon lists at $2 per million input tokens, more than double Hy4. At the OpenRouter half price, Hy4 undercuts Argon by a wide margin, which is the main reason to try it now rather than later.

How to Try It

The fastest route for most readers outside China is OpenRouter, where the model ID is tencent/hy4 and the API is OpenAI-compatible, so existing code needs only a base URL and model name change. You need an OpenRouter account and API key, and usage is billed per token at the rates above. Usage data on the OpenRouter page shows the biggest traffic so far comes from coding agents: Claude Code alone accounts for 88.6 billion tokens, followed by the Strix security agent and Nous Research's Hermes Agent.

Inside Tencent's own products, Hy4 is already the model behind WorkBuddy, CodeBuddy, Yuanbao, and WeCom's Dayuan assistant. WorkBuddy is the most interesting of these for office work: Tencent's evaluation gives Hy4 an average score of 83.18 and a 96.15% pass rate across its WorkBuddy scenarios, ranking first in nine of twelve scenarios in the mainland Chinese market. Those are Tencent's numbers on Tencent's product, but they point to where the model is strongest today: documents, presentations, and multi-step office tasks rather than raw coding.

What you cannot do yet is download it. The weights are not on Hugging Face as of November 8; the repository page for the full model is not public, and only the August preview is available under the Apache 2.0 licence. If you need a model you can self-host today, the preview is the only option, and it is text-only, without the vision and computer-use capabilities that define the full release.

The Open-Weights Question

This is the part of the launch that deserves the most scrutiny. Tencent's press release says it "plans to release model weights and system cards in the coming days," and its research page says it has partnered with the vLLM and SGLang communities and Inferact Inc. to support the open-source ecosystem, with a full system card including safety evaluations to follow. No date has been given, and four days after launch there is still no public repository for the full model.

That matters because "open-weight" is doing a lot of work in the coverage of this release. Right now Hy4 is a hosted API with a promise attached, not an open model you can inspect or run. The preview's Apache 2.0 licence is a genuine precedent, and Tencent did open-source two science models built with Hy4, HyCell for single-cell biology and HyPDE for physics fields, on the same day. But until the full weights land, the honest label is "planned open-weight," and anyone choosing Hy4 over a rival on the strength of openness should wait for the download link.

Safety, Hallucinations, and the Fine Print

Tencent built safety into the announcement, and some of it is substantive. Hy4 was trained on defensive security data only: proof-of-concept development and vulnerability remediation are in, while offensive material related to exploitation is explicitly excluded from training. Security teams from Tencent's Xuanwu Lab and Keen Security Lab assessed the model across cryptographic reverse engineering, binary deobfuscation, and vulnerability exploitation, and reported no significant concerns in the areas evaluated. On biosafety, Hy4 scores 30.4 out of 100 on Tencent's own Hy-BioExperimentBench, which the research page describes as below the high-risk threshold, though its open-ended experimental design ability is weaker than frontier models.

The hallucination numbers need a warning of their own. Tencent's research page says its self-built test cut the hallucination rate from 15.76% on the preview to 5.45%, with errors from unsupported assumptions falling from 23.1% to 5.6%. The corporate press release gives different figures for the same improvement: 12.95% down to 7.2%. These are two different internal tests reported as if they were one story, so do not quote either pair as the definitive hallucination rate. What both versions agree on is the direction: hallucinations roughly halved from the preview, on Tencent's own tests.

A few more details worth knowing before you commit. The model is positioned as groundwork for recursive self-improvement, with Tencent's Hyra 2.0 harness used to optimize pre-training infrastructure for smaller models on the same architecture. The preview was already the fourth most-used model in OpenRouter's weekly ranking at 9.48 trillion tokens, according to launch-week coverage, which suggests real developer demand rather than launch-day curiosity. And the 50% TokenHub discount ends December 31, 2026, while the free WorkBuddy and CodeBuddy period ends November 15, so the cheapest window to try it is this week.

Common Mistakes to Avoid

Do not call Hy4 open source yet. It is a hosted model with a published plan to release weights, and the two are not the same thing. Do not quote a single hallucination rate, since Tencent's own pages disagree with each other. Do not assume the OpenRouter half price is permanent, since the preview still lists at full price on the same platform. Do not judge the model on the finance or WorkBuddy evaluations alone, since both were run by Tencent on its own products. And do not expect it to match Claude Opus 5.5 or GPT-6.1 Sol on hard maths or science: Tencent's own limitations chart puts the gap at tens of points, not a few.

FAQ

Is Hy4 open source?

Not yet. The August preview's weights are public under the Apache 2.0 licence, but the full November 4 release's weights have not been published as of November 8, 2026. Tencent says it plans to release them in the coming days with vLLM and SGLang support.

How much does 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. OpenRouter and Tencent Cloud TokenHub both currently charge half that, and WorkBuddy and CodeBuddy are free until November 15, 2026.

Can Hy4 use a computer?

Yes. It can operate command-line and graphical interfaces, with specialized training on Blender, CAD tools, and game engines. On OSWorld 2.1 it scores 78.9, close to Claude Opus 5.5 at 81.8, though that figure comes from Tencent's own benchmark chart.

How does Hy4 compare to Claude Opus 5.5?

It is close on computer use (78.9 vs 81.8 on OSWorld 2.1) and on agent exams (41.0 vs 43.8), but far behind on coding (44.4 vs 66.4 on Terminal Bench 4.0) and on hard maths, where Tencent's own test has Hy4 at 13.0 against Claude at 91.3. All figures are Tencent-reported.

When will the Hy4 weights be released?

Tencent has not given a date, only "in the coming days" as of the November 4 announcement. The full model's Hugging Face repository was still not public on November 8, 2026.

Where can I try Hy4?

Through OpenRouter as tencent/hy4, on Tencent Cloud TokenHub, or inside Tencent's WorkBuddy, CodeBuddy, Yuanbao, and WeCom Dayuan apps. It is not available through OpenRouter's EU in-region routing.

Hy4 is a real step forward for Tencent and a useful new option for anyone who needs a model that can see a screen and drive it. Its computer-use scores are close enough to the frontier to matter, its price undercuts the premium Western models, and the free WorkBuddy window makes this week the cheapest time to try it. The two things to watch are the ones Tencent has not delivered yet: the open weights, and independent benchmarks. Until both arrive, Hy4 is best described as a strong hosted model with an open promise attached, not the open-weight flagship the headlines suggest. Try it on a real office or GUI task this week while it is half price, and judge it on that, not on the launch chart.