Analysis
Tencent releases Hy4, a 770B-parameter model with vision and computer use, but says the final open weights are still to come, and the benchmarks are its own
Tencent released Hy4, a 770B model with vision and computer use. Only the preview's weights are open; the final ones are promised. Benchmarks are its own.
On November 4, 2026, Tencent released Hy4, a mixture-of-experts model with 770 billion parameters in total, the first of its Hy series with vision and computer use. The open weights are still a promise: only the August preview's weights are published. The benchmarks are Tencent's own. This is neither investment, legal nor purchase advice.
The essentials
- Release: November 4, after a preview on August 28; vision and computer use are new (explainx).
- Open?: only the preview's weights are published, under Apache 2.0. For the full model, Tencent says it plans to publish them, with no date or announced license.
- Benchmarks: Tencent's own; Hy4 is first on none of 12 tests (explainx).
- Not established: date and license of the final weights, performance outside Tencent's tests.
What does Tencent say?
Hy4 has 770 billion parameters in total (MoE) and a context of about 1 million tokens. The preview's spec sheet gives 49 billion active parameters per token (eesel, MadRobot); explainx gives no active count for the final model. It reads charts, documents and photos; in a demonstration it spots a roof section of about 2 m in a 2.8-billion-pixel aerial photo (MadRobot). For computer use (command line and graphical interfaces) it was trained on Blender, CAD tools and multi-application tasks (explainx). Tencent also says it removed exploitation and penetration-testing data from training and that its labs see no marked offensive cyber risk; a full system card is announced (MadRobot, per Tencent). Performance outside Tencent has not been tested (DEV Community).
"Open": the preview versus the full model
The preview weights are on Hugging Face under Apache 2.0 (eesel, MadRobot, explainx). For the full Hy4, Tencent says it plans to publish the weights, with vLLM and SGLang support; no date is given and the license is not announced (MadRobot, explainx, DEV Community). The preview weights are published; for the full model, the listed sources describe a planned release, with no date. So "open" is true today only for the preview. A smaller Hy4 Flash is announced as coming soon, with no date, price or specifications. For the wider debate, see our article on regulating open-weight models.
Source: MadRobot, explainx, eesel.
What do Tencent's benchmarks show?
The tests are Tencent's, re-run by Tencent for seven models; they are not independent. On Terminal Bench 4.0, Hy4 scores 44.4 against 66.4 for Claude Opus 5.5, 55.5 for GPT-6.1 Sol and 66.8 for Kimi K3.1 Pro. On OSWorld 2.1: 78.9 against 81.8, 75.2 and 79.6. On ProgramBench: 30.0 against 56.5 for Claude Opus 5.5. Across 12 Tencent tests, Hy4 ranks first on none (explainx). Tencent acknowledges a significant gap with the best closed models, notably on frontier math and long software projects. In its financial test (213 experts, 579 tasks), Hy4 scores 3.04 out of 4 against 2.97 for DeepSeek.
Source: Tencent via MadRobot, explainx.
Availability and prices
Hosted access: Tencent Cloud (TokenHub), OpenRouter and Tencent apps (Yuanbao, WorkBuddy, CodeBuddy, ima). Tencent Cloud charges 6 yuan per million input tokens and 18 for output, half price on TokenHub until December 31, 2026 (MadRobot, explainx). OpenRouter shows $0.42 and $1.25 per million (MadRobot), while preview coverage cited $0.834 and $2.501 (eesel, DEV Community): the figures diverge, so check the price page.
What is not established?
The date and license of the final weights; whether the full Hy4 is open in the preview's sense; the active parameters of the final model; performance outside Tencent's tests; how well computer use works; current OpenRouter prices; the date of Hy4 Flash.
Comparison table
| Point | What the sources say | Status |
|---|---|---|
| Preview weights | Published, Apache 2.0 | Reported |
| Full Hy4 weights | Planned, no date or license (Tencent) | Announced by Tencent |
| Size | 770B total, about 1M context | Announced by Tencent |
| Benchmarks | First on none of 12 tests (explainx) | Announced by Tencent |
| Real-world computer use | Not tested outside Tencent | Not established |
Frequently asked questions
Are Hy4's weights published?
Only the preview's. Tencent says it plans to publish the full model's, with no date or license announced.
Does Hy4 beat Claude or GPT?
Not in Tencent's own tests shown here: it trails Claude Opus 5.5 on all three.
Does computer use really work?
Not verified outside Tencent's tests.
Method and limits
Sources as of November 5, 2026. Figures are Tencent's, relayed by the outlets listed and attributed to each. We tested nothing and give no advice.
Sources
Spotted an error in this article? Report it · Our corrections policy
