What is Hy4?
Hy4 is the full release of Tencent's flagship Hy model (formerly Hunyuan), launched on 4 November 2026 as the successor to Hy4 preview from August. It is the first Hy model that reads images and the first that can operate a computer, through both the command line and on-screen interfaces; Tencent says it trained it on professional software such as Blender and CAD tools. It also claims better instruction following, long-document reading and long multi-step tasks than the preview.
Tencent's launch post gives no parameter count. The preview it replaces is a 770-billion-parameter mixture-of-experts model with 49 billion active per token, and OpenRouter lists the same 1M-token context, with up to 65,536 output tokens.
Tencent calls Hy4 open source, but at launch the weights were not out: it says weights and system cards will follow "in the coming days" and names no licence. Until then it is an API model, on Tencent Cloud TokenHub, OpenRouter and Tencent apps including Yuanbao, WorkBuddy and CodeBuddy. A cheaper Hy4 Flash is promised next.
How much does Hy4 cost?
| Plan | Price | What you get |
|---|---|---|
| OpenRouter (Tencent) | $0.834 in / $2.501 out per 1M tokens | Cached input $0.042; 50% launch discount until 31 December 2026 |
| Tencent Cloud TokenHub | About $0.90 in / $2.70 out per 1M tokens | 50% off until 31 December 2026 |
| Weights | Not yet released | Promised with system cards; licence not named |
OpenRouter (Tencent) list price $0.834 input / $2.501 output per 1M tokens, cached input $0.042; Tencent quotes about $0.90 / $2.70. TokenHub is 50% off until 31 December 2026. Weights not yet released.
Compare this price with other models
Is Hy4 worth it?
Hy4's case is price, not rank. On Artificial Analysis's Intelligence Index it scores 35, level with Grok 4.7 and second among Chinese labs behind Kimi K3.1 Pro at 46, but below the median of 40 for models in its price band. It costs about $1.03 per task to run that index, roughly a tenth of what Grok 4.7 costs for the same score.
Tencent's own numbers point the same way. In its comparison of seven models across 12 benchmarks, as read from its launch chart by third parties, Hy4 comes first on none. It scores 78.9 on OSWorld 2.1, the computer-use test, against 81.8 for Claude Opus 5.5 and 79.6 for Kimi K3.1 Pro, and 44.4 on Terminal-Bench 4.0 against Kimi's 66.8. Its best showing is a blind finance test Tencent ran itself: 213 professionals rated it 3.04 out of 4, ahead of DeepSeek-V4.1-Flash at 2.97 and Kimi K3.1 Pro at 2.68.
The catch is reliability. Tencent says its internal hallucination rate fell from 12.95% in the preview to 7.2%, yet Artificial Analysis's AA-Omniscience test finds that when Hy4 does not know an answer it invents one 89% of the time. It is also wordy: it spent 160 million output tokens on the index, more than twice the median, which eats into the low per-token price.
Treat the open-source label as pending. The preview's weights are Apache 2.0 on Hugging Face; the full model's are not public, with no date or licence named. For teams that want an inexpensive agent model with screen control today, Hy4 is worth a trial on their own tasks; for self-hosting, the preview remains the only Hy4 you can download.
Where it earns its place
An operations team automating desktop chores in Chinese office software
Have an agent fill forms, move files and drive GUI apps without APIs
Hy4's computer use combines command line and screen control and scores 78.9 on OSWorld 2.1 in Tencent's runs, a few points behind Kimi K3.1 Pro. Inside WorkBuddy it can read WeCom chats, email and calendars.
A fintech research desk at a mid-sized brokerage
Draft fund and securities analysis from filings and charts
Tencent built thousands of finance workspaces for training and reports a 3.04 out of 4 rating from 213 finance professionals. Check every figure: AA-Omniscience shows a high rate of invented answers.
A startup already running Hy4 preview
Add image input to an existing agent without changing the bill
Switch the model id from tencent/hy4-preview to tencent/hy4 on OpenRouter: same $0.834 / $2.501 list price, plus image input. Keep the preview if you need weights you can host.
Before you commit
- Setup
- Call tencent/hy4 on OpenRouter or through Tencent Cloud TokenHub; thinking is on by default, so set an output cap. WorkBuddy and CodeBuddy offered it free from 4 to 15 November.
- Integrations
- OpenAI-style API on OpenRouter (Tencent is the only provider listed); TokenHub; Yuanbao, WorkBuddy, CodeBuddy, ima and WeCom's assistant. Tencent says API access on Hermes and OpenCode is rolling out.
- Data and privacy
- API only for now: prompts go to Tencent Cloud, which does not state retention publicly. No weights, so no self-hosting. Tencent says its security training excluded offensive exploitation material.
- Cost at scale
- At list price, 1,000 agent tasks of 200K input and 20K output tokens cost about $217 before caching, half that on TokenHub until 31 December. Artificial Analysis puts its index runs at about $1.03 per task.
- Limits
- 1M-token context, 65,536-token output cap on OpenRouter. High hallucination rate on AA-Omniscience; long, verbose reasoning; Tencent says open-ended experimental design trails frontier models.
What’s good
- First Hy model with image input and computer use
- Same price as the preview, with 50% off on TokenHub until 31 December
- Third-party scored: Artificial Analysis gives it 35
- 1M-token context and 65,536 output tokens on OpenRouter
What’s not
- Leads none of Tencent's 12 launch benchmarks
- AA-Omniscience: invents an answer 89% of the time it does not know
- Weights and licence not yet published
- Verbose: more than twice the median output tokens on the index
Best for
Teams that want a low-cost agent model with screen control and image input, and can wait for weights before self-hosting.
Hy4 alternatives
Hy4: common questions
- Is Hy4 worth it?
- Hy4 is a cheap, capable second tier: vision and computer use at preview prices, but it leads none of Tencent's own benchmarks and its open weights are still a promise. We rate it 3.8 out of 5.
- How much does Hy4 cost?
- OpenRouter (Tencent) list price $0.834 input / $2.501 output per 1M tokens, cached input $0.042; Tencent quotes about $0.90 / $2.70. TokenHub is 50% off until 31 December 2026. Weights not yet released.
- What are the downsides of Hy4?
- Leads none of Tencent's 12 launch benchmarks; aA-Omniscience: invents an answer 89% of the time it does not know; weights and licence not yet published; verbose: more than twice the median output tokens on the index.
- Who is Hy4 best for?
- Teams that want a low-cost agent model with screen control and image input, and can wait for weights before self-hosting.
- How much does Hy4 cost for a team or in production?
- At list price, 1,000 agent tasks of 200K input and 20K output tokens cost about $217 before caching, half that on TokenHub until 31 December. Artificial Analysis puts its index runs at about $1.03 per task.
- Is Hy4 safe for company data?
- API only for now: prompts go to Tencent Cloud, which does not state retention publicly. No weights, so no self-hosting. Tencent says its security training excluded offensive exploitation material.
- What are the best Hy4 alternatives?
- The closest alternatives we have reviewed are Hy4 preview, Kimi K3.1 Pro, DeepSeek-V4.1-Flash and GLM-5.3.
Answers checked Nov 11, 2026.