On November 4, 2026, Tencent Hunyuan released the official version of Hy4, a mixture-of-experts (MoE) model with 770B total parameters and a 1M context window. It’s the first Hy-series model with vision capabilities and the first to support computer use through both the command line and a graphical interface. The API list price is RMB 6 per million input tokens, RMB 18 per million output tokens, and RMB 0.3 per million cached tokens. Calls through Tencent Cloud TokenHub are 50% off through December 31, 2026.
This isn’t a benchmark victory lap. Tencent says Hy has “joined the ranks of the strongest open-source models for the first time,” but we’ll treat that as a marketing claim. Instead, we’ll look at Hy4’s actual rankings across the 12 benchmarks in the main chart, external feedback from launch day, and the shortcomings Tencent itself acknowledges. That should help you decide which tasks it’s a good fit for.
What you’ll get: By the end, you’ll know which tasks are worth trying Hy4 on right away, where sticking with your current model makes more sense, and which channel to use—and roughly what it’ll cost—if you want to try it today.
Hy4 at a Glance
| Item | Details | Source |
|---|---|---|
| Release date | November 4, 2026 (official release) | Tencent Hunyuan release page |
| Architecture and parameters | MoE, 770B total parameters; the preview model card lists 49B active parameters | Tencent News / Hugging Face preview model card |
| Context window | 1M (OpenRouter lists 1,048,576 tokens) | OpenRouter model page |
| Maximum output | 64K (based on preview documentation; official release documentation hasn’t been updated yet) | TokenHub Hunyuan invocation guide |
| New capabilities | First-time support for vision and computer operation (Computer Use) | Tencent Hunyuan release page |
| API protocols | Compatible with OpenAI Chat Completions, OpenAI Responses, and Anthropic Messages (based on preview documentation) | TokenHub documentation |
| Weights | “Planned for open source”; preview weights are available under Apache 2.0, but official release weights haven’t been released | Tencent Hunyuan / Hugging Face |
Compared with Hy4 preview, released on August 28, Hy4 scales up both pretraining and post-training. The main improvements are in instruction following, long-document comprehension, long-horizon tasks, and deep research. According to the preview documentation, thinking mode is enabled by default, and the API returns a reasoning_content field. You’ll need to account for processing that content when integrating it.
Hy4 Vision and Computer Tasks: Seeing Images and Operating Software for the First Time
For vision, the official announcement highlights fine-grained image processing, complex chart analysis, and visual mathematical reasoning. One standout demo locates a rooftop feature about 2 meters across in a 2.8-billion-pixel aerial image. Another reassembles a 900-piece puzzle to recreate Monet’s Bordighera.
Computer operation is the more noteworthy change. For the first time, Hy4 supports workflows that combine the command line with graphical interfaces. It was also trained specifically for professional software such as Blender and CAD tools, as well as long-running tasks across applications. According to the Tencent Cloud Developer Community, the training team worked with the Tencent Cloud Agent Runtime team to build a sandbox with hundreds of applications installed for executing tasks and verifying results.
Hy4 Frontend and Game Development: Support for UE, Unity, and Godot
On the frontend, the stated focus is on complex interactions, business logic, state management, and frontend-backend integration. The goal is interfaces that “not only look good but are fully functional.” One demo is a single-file HTML “Tencent Penguin Island 3D Digital Sandbox” with Three.js inlined, so it doesn’t depend on a CDN.
Tencent IEG game developers contributed to the game development work, with deep integration for UE, Unity, and Godot. Hy4 is also beginning to cover multiplayer online game development. Compared with Hy4 preview, GameDev improved by 19.1% and GameLogic by 9.8%. Note that both figures compare Hy4 with its own previous version, not with other models.
Hy4 in Finance and Office Work: A Significant Drop in Hallucination Rate
The most complete data comes from financial use cases. Tencent Financial Technology (FiT) organized a crowdsourced evaluation with 213 experts, including 153 from leading external financial institutions, covering 579 real-world financial tasks. On a four-point scale, Hy4 scored 3.04, DeepSeek-V4.1-Flash 2.97, Kimi K3.1 Pro 2.68, and GLM-5.3 2.68. Hy4’s lead over DeepSeek was modest.
The hallucination rate is arguably more meaningful in practice. In Tencent’s internal search tests, Hy4’s hallucination rate fell from 15.76% in the preview version to 5.45%, while the error rate caused by guessing dropped from 23.1% to 5.6%. For office work, people and the model can collaboratively edit Word, Excel, and PowerPoint files in real time. Partner Manus says the completion rate for office document generation and editing tasks is close to 97%, though that’s a vendor-reported figure.
Hy4 in Research and Personal Assistance: From Research Tasks to WeCom
In Tencent’s internal TenureTrail evaluation, which covers 123 executable research tasks across seven disciplines, Hy4 ranked first among open-source models. One official example involved reconstructing a mouse embryo in 3D from 84 unordered 2D slices, some of which were mirrored, containing 6.99 million cells. Working offline and without a dedicated registration tool, Hy4 aligned the slices and completed the 3D reconstruction in 1.5 hours, with no human intervention.
As a personal assistant, Hy4 connects to chat history, email, documents, calendars, and meetings through WorkBuddy’s WeCom suite. It can piece together the progress history of multiple projects and flag follow-ups and upcoming deadlines. Hy4-powered team collaboration features will launch on WorkBuddy alongside the updated “Projects” feature.
How to Read Hy4’s Benchmarks: Mostly 3rd or 4th Place
First, a note on methodology. All the data below comes from Tencent’s own tests. Tencent reran the competing models under consistent settings, so these aren’t scores published by the respective vendors and shouldn’t be treated as an independent cross-vendor leaderboard. Third-party independent evaluations currently cover only Hy4 preview, so they can’t be taken as representative of the final release.
| Benchmark | Hy4 | Highest score | Hy4 rank (7 models) | Open-source comparison: Hy4 vs Kimi K3.1 Pro |
|---|---|---|---|---|
| ProgramBench | 30.0 | Claude Opus 5.5 56.5 | 3rd | Hy4 wins (24.5) |
| FrontierCode 2.0 | 31.0 | Claude 46.4 | 3rd | Hy4 wins (25.5) |
| 3DArena-Modeling | 37.1 | Claude 43.4 | 3rd | Hy4 wins (36.1) |
| Terminal Bench 4.0 | 44.4 | Kimi 66.8 | 4th | Kimi wins |
| Agents’ Last Exam | 41.0 | Kimi / Claude 43.4 | 3rd | Kimi wins |
| AutomationBench | 56.2 | Kimi 65.6 | 3rd | Kimi wins |
| E-Bench 2.0 | 70.0 | Claude 80.0 | 2nd | Hy4 wins (50.0) |
| Hy-BrowseComp 2.0 | 59.9 | GPT-6.1 Sol 61.7 | 3rd | Hy4 wins (59.3) |
| PostTrainBench | 35.2 | Claude 44.2 | 4th | Kimi wins |
| Terminal Bench Science | 19.8 | GPT 60.9 | 4th | Kimi wins |
| TenureTrail | 53.6 | Claude 59.0 | 3rd | Hy4 wins (51.3) |
| OSWorld 2.1 | 78.9 | Claude 81.8 | 3rd | Kimi wins (79.6) |
The takeaway is clear: Hy4 doesn’t rank first overall on any of the 12 benchmarks. Its best result is second place on E-Bench 2.0, followed by third on OSWorld. Among the open-source models, it ties Kimi K3.1 Pro at 6–6. Hy4’s strengths are concentrated in code generation, 3D modeling, retrieval, and research tasks. Kimi, meanwhile, has a clear lead in terminal operations, automation, and scientific terminal tasks, with a gap of more than 2× on Terminal Bench Science.
Hy4 trails Claude Opus 5.5 and GPT-6.1 Sol on most benchmarks. The gaps are even wider in the supplementary results: on Hy-FrontierMath T5, Hy4 scores just 13.0 versus GPT-6.1 Sol’s 97.5; on Goal2Repo, Hy4 scores 23.2 versus Claude’s 74.0. On Foundation Model Bench, Hy4’s 28.1 is also below DeepSeek’s 36.3.
Hy4 Day-One Testing: Feedback From a Small Sample
The OpenRouter page shows that Hy4 has a P50 throughput of around 63 tok/s, a P50 time to first token of around 2.9 seconds, and 99.87% availability over three days. X user @leo114119 measured throughput of 60–70 tok/s on WorkBuddy and called it “not fast.” Another assessment was that “real-world performance falls short of the benchmark scores.” @VeknQAQ described the results as “so-so.”
These are all individual impressions from November 4, and the sample is too small to draw conclusions. Still, they’re a reminder that official benchmark scores and real-world experience can differ. Before deploying Hy4, test it against your own set of tasks.
Hy4 Pricing and Where to Access It
| Billing item (per million tokens) | TokenHub list price | TokenHub 50% off | OpenRouter list price | OpenRouter 50% off |
|---|---|---|---|---|
| Input | 6 yuan | 3 yuan | $0.834 | $0.417 |
| Output | 18 yuan | 9 yuan | $2.501 | $1.2505 |
| Cache hit | 0.3 yuan | 0.15 yuan | $0.042 | $0.021 |
TokenHub’s discount runs through December 31, 2026, at 23:59:59 Beijing time. Hy4 and the preview have the same pricing. At an estimated exchange rate of 7 yuan per US dollar, OpenRouter’s list prices work out to roughly 5.8 yuan for input, 17.5 yuan for output, and 0.29 yuan for cache hits—about the same as TokenHub. OpenRouter shows a cache hit rate of around 93%, so tasks with long contexts and frequently repeated prefixes may cost less than the list price suggests.
| Channel | Access method | Notes |
|---|---|---|
| Tencent Cloud TokenHub | API, model ID hy4 |
Preview is hy4-preview |
| OpenRouter | API, model ID tencent/hy4 |
Currently showing 50% off |
| WorkBuddy / CodeBuddy | Use directly in the product | Limited free quota from November 4–15 |
| Yuanbao, WeCom Dayuan, ima | Consumer or workplace products | No development required |
| Hermes, OpenCode | API integration already available | Suited to coding agent use cases |
💡 Getting started: Hy4 isn’t currently among APIYI’s 277 online models (snapshot from 2026-11-04 10:00). Check the platform’s model list for current availability and pricing; we’ll add this model as soon as possible. If you’d like to establish a baseline for comparison, you can call the already available
deepseek-v4.1-flashthrough APIYI at apiyi.com. It costs $0.3 for input, $1.2 for output, and $0.006 for cache reads per million tokens. The interface is OpenAI-compatible; just setbase_urltohttps://api.apiyi.com/v1.
Hy4’s Limitations: Gaps Acknowledged by Tencent
On its release page, Tencent acknowledges that Hy4 still lags significantly behind leading closed-source models, particularly in frontier mathematics, long-running development tasks, model iteration, and visual question answering. Its BuddyVQA Atlas score supports that assessment: Hy4 scored 55.6, compared with Claude’s 75.3. The cheaper Hy4 flash is only described as “coming soon,” with no release date. The maximum output of 64K comes from the preview documentation; the documentation for the final release hasn’t been updated yet. On safety, Hy4 scored 30.4/100 in its Hy-BioExperimentBench biosafety self-assessment. Tencent says this is below the high-risk threshold, and the final release’s model system card hasn’t been published yet.
Hy4 Impact Analysis
Hy4’s Impact on Developers: Which Tasks Are Worth Switching For?
| Task Type | Recommendation | Why |
|---|---|---|
| Financial research reports and investment Q&A | Try Hy4 first | Scored 3.04 in expert testing, with the hallucination rate down to 5.45% |
| Front-end pages and 3D / game prototypes | Worth trying | Leads open-source models on 3DArena and FrontierCode |
| Desktop software operation agents | Evaluate alongside Kimi | Scores 78.9 on OSWorld, still below Kimi’s 79.6 |
| Terminal automation and scientific computing | Stick with your current model | Trails Kimi significantly on the Terminal Bench series |
| Frontier mathematics and long-running development tasks | Stick with top closed-source models | Scores just 13.0 on FrontierMath T5 |
| Cost-sensitive batch tasks | Compare prices first | Use DeepSeek V4.1 Flash through APIYI at apiyi.com as a cost baseline |
For teams choosing a model, Hy4’s biggest change is that the open-source camp now has a flagship model covering both vision and computer operation. If your business needs to interpret screenshots or operate Blender or CAD, you previously had to turn to closed-source models or stitch together multiple models. Now there’s a potential candidate for private deployment—provided the final model weights are actually released. For now, Hy4 is available only through the API and Tencent’s own products.
Hy4’s Impact on the Industry: A Fourth Player Enters the Open-Source Flagship Race
After DeepSeek, Kimi, and GLM, Hy4 makes Tencent the fourth Chinese company to bring an open-source flagship to this level. Its 6:6 result against Kimi suggests there’s no longer a single undisputed leader among the top open-source models. An upgrade at the same price as the preview, combined with a 50% discount, will put further pressure on Chinese API prices. For content teams and small and midsize businesses, no-code options like Yuanbao, ima, and WeCom Dayuan may bring Hy4 to users sooner than the API.
Hy4 FAQ
Q1: Is Hy4 open source? Can I download the weights now?
The official statement says the team “plans to open-source the model weights.” Hy4 preview weights are available on Hugging Face under Apache 2.0. As of 2026-11-04, the full Hy4 release weights haven’t been released.
Q2: Can Hy4’s benchmark scores be compared directly with results published by other vendors?
Not directly. Tencent ran all the tests itself under a consistent setup, and third-party evaluations have only covered the preview so far. Your safest bet is to test with your own business data. For example, you can first run DeepSeek V4.1 Flash on APIYI apiyi.com to establish a baseline, then compare its results with Hy4’s output.
Q3: What’s the cheapest way to access Hy4 right now?
Through December 31, Hy4 is available at 50% off on Tencent Cloud TokenHub: 3 yuan for input, 9 yuan for output, and 0.15 yuan for cache. If you just want to try it, WorkBuddy / CodeBuddy offers limited free usage from November 4–15.
Q4: Can I access Hy4 through APIYI?
As of the 2026-11-04 10:00 snapshot, Hy4 isn’t yet on APIYI’s list of available models. Check the platform’s model list for current availability and pricing; we’ll roll out the model as soon as possible. In the meantime, you can use the same API key on APIYI apiyi.com to access available models such as `deepseek-v4.1-flash` and set up your evaluation workflow.
Hy4 Key Takeaways
- Specs: 770B MoE, a 1M context window, and support for vision and computer use for the first time
- Benchmark rankings: No first-place finishes across the 12 benchmarks in the main chart; its best result was second place on E-Bench 2.0
- Open-source comparison: Tied Kimi K3.1 Pro at 6:6, with each model excelling at different tasks
- Most practical highlight: Scored 3.04 in crowdsourced financial-task testing, with the hallucination rate falling from 15.76% to 5.45%
- Pricing: RMB 6 / 18 / 0.3 per million tokens, with 50% off on TokenHub through the end of the year
- First-day feedback: Throughput of around 63 tok/s; some users reported that real-world performance didn’t match the benchmark scores
- Getting started: Call the model through TokenHub or OpenRouter; use models already available on APIYI apiyi.com to build a comparison baseline first
Summary
Here’s the bottom line on Hy4:
- Capabilities: Vision and computer use address gaps in its open-source flagship capabilities, with the strongest improvements in finance and office workflows.
- Positioning: It’s in the top tier of open-source models, but it still trails Claude Opus 5.5 and GPT-6.1 Sol by a noticeable margin, especially in frontier math and long-horizon development.
- Cost: It’s priced the same as the preview, and the 50% discount makes it competitive among flagship models from China.
If you work on financial research reports, office documents, or 3D / frontend prototypes, it’s worth testing Hy4 on real tasks while the discount lasts. For terminal automation and difficult reasoning tasks, there’s no need to rush into switching yet. When building your evaluation baseline, you can access comparison models already available through APIYI apiyi.com in one place.
Suggested Locations for Screenshots
- Tencent Hunyuan Hy4 launch page, above the fold: Place after the cover image at the beginning of the article. Caption: Source: Tencent Hunyuan, hy.tencent.com/research/hy4. Captured on 2026-11-04.
- Tencent Cloud TokenHub model pricing table: Place after the first table in “Hy4 Pricing and Availability.” Caption: Source: Tencent Cloud TokenHub model pricing documentation. Captured on 2026-11-04.
- OpenRouter Hy4 model page: Place in the “Hy4 First-Day Hands-On Test” section. Caption: Source: OpenRouter tencent/hy4 model page. Captured on 2026-11-04. Throughput and latency data may change over time.
📚 References
- Tencent Hunyuan, “Hy4 Official Launch”: Official launch page and three benchmark charts
- Link:
hy.tencent.com/research/hy4 -
Notes: Source for specifications, capability demonstrations, the main chart covering 12 benchmarks, and appendix data
-
Tencent Cloud TokenHub model pricing: Hy4 list prices and limited-time discounts
- Link:
cloud.tencent.com/document/product/1823/130055 -
Notes: RMB pricing and the end date for the 50% discount
-
Tencent Cloud TokenHub Hunyuan invocation guide: API protocol and context parameters
- Link:
cloud.tencent.com/document/product/1823/132252 -
Notes: Maximum input and output limits and thinking mode details for the preview version
-
OpenRouter Hy4 model page: USD pricing and performance statistics
- Link:
openrouter.ai/tencent/hy4 -
Notes: Throughput, time to first token, and availability
-
Hugging Face Hy4-preview model card: Preview architecture details and weights
- Link:
huggingface.co/tencent/Hy4-preview -
Notes: Active parameters, layer count, number of experts, and more
-
APIYI model pricing overview: Available models and pricing
- Link:
docs.apiyi.com/models - Notes: Snapshot from 2026-11-04 at 10:00, with 277 models available
Author: APIYI Technical Team
Technical discussions: Contact us through the APIYI homepage at apiyi.com. For more integration resources, visit the documentation center at docs.apiyi.com.
