Tencent releases full Hy4 AI model that can see images and use a computer

Tencent released the full version of Hy4 on November 4, 2026, ten weeks after the August preview covered here. Visual understanding and computer use are the two main additions. The preview arrived with open weights and a model card on day one, but this release does not. Tencent says it “plans to release model weights and system cards in the coming days”. Hy4 can now read images and charts, and it can operate command-line and graphical interfaces, both for the first time.

Tencent describes the launch as its first entry into the top group of open models. It says Hy4 will “break into the top tier of open source AI models for the first time”. Almost all of the evidence for that claim is Tencent’s own.

What changed since the Hy4 preview

The preview was a text-only model. According to Tencent, the full Hy4 adds three things:

• Visual reasoning. It handles fine-grained image processing, complex chart analysis and mathematical reasoning over visuals, combining perception with tool calls.
• Computer use. It can interact with both command-line and graphical user interfaces. Tencent says it received specialised training on professional software such as Blender and CAD.
• Better core behaviour. Tencent reports improvements in instruction following, long-context comprehension and long-horizon agentic tasks, meaning multi-step jobs that run over many actions.

OrcaRouter, an API routing service, reviewed the release. It says Tencent also claims a shorter chain of thought, the step-by-step reasoning a model writes out before it answers. That matters because the preview’s own model card admitted two problems: the model reasoned longer than necessary on hard tasks and over-verified its own answers. SiliconFlow’s guide to the preview flagged the same limitations. OrcaRouter notes that Tencent published no before-and-after token counts, so the efficiency gain cannot yet be measured.

Tencent’s demonstration of the vision feature uses a 2.8-gigapixel aerial photograph. In it, the model finds a two-metre roof component. OrcaRouter calls this a vendor illustration with no published visual benchmark attached.

Hy4 specs: what is confirmed and what is carried over

The size figures need care. OrcaRouter reports that the launch announcement does not restate the parameter count or context length. It says the figures in circulation come from the preview. The News International does report that Hy4 “carries 770 billion parameters, up from 295 billion in Hy3”. Tencent has also not described the vision encoder, the separate component that turns images into input a language model can use. That includes its size, its training data and whether it is a separate model.

Spec	Hy4 Preview (August)	Hy4 (November)
Total parameters	770B	770B reported; not restated by Tencent per OrcaRouter
Active parameters per token	49B	Preview figure in circulation
Architecture	Mixture-of-experts: 256 routed experts plus 1 shared per layer, 8 routed active per token	No new description published
Context window	1M tokens	1,048,576 tokens, preview figure
Input	Text	Text and images
Computer use	No; built for tool schemas	Command line and GUI
Licence	Apache 2.0	Unstated
Model card	Published with weights	Not yet; system cards promised

A mixture-of-experts (MoE) model stores many specialised sub-networks, called experts, and sends each token through only a few of them. That is why only 49 billion of the 770 billion parameters do work on any given token. SiliconFlow notes that hosting the model still requires storing the full checkpoint.

Where Hy4 is available and what it costs

The model is live through API providers now, but the downloadable files are not. A ZenMux listing for the TencentCloud provider shows an entry for “tencent/hy4” at $0.834 per million input tokens and $2.501 per million output tokens. Those are the same figures the listing shows for Hy4 Preview. OrcaRouter likewise reports that the price did not move. It describes the two models as “the same family, the same size and, as of today, the same price”.

Tencent built the release around its own workplace software. Hy4 powers a broad update to WorkBuddy, Tencent’s workplace AI tool. Through WorkBuddy it integrates with WeCom, Tencent’s workplace messaging app, and reaches chats, emails, documents, calendars and meetings. Tencent also says new team collaboration features will arrive with an upcoming WorkBuddy feature called “Projects”. For computer use, OrcaRouter reports that Tencent runs the model in a sandbox built on Tencent Cloud’s Agent Runtime, with several hundred applications installed.

Nikkei Asia describes the broader strategy: Tencent has moved from proprietary to open-source models to attract customers to its related cloud services.

Tencent’s benchmarks, and how much they show

The full release has no benchmark appendix. It publishes no Terminal-Bench or SWE-bench results, two widely used public coding tests. Instead, Tencent reports results from its own task suites and expert panels:

• Finance panel. Tencent ran a blind evaluation with Tencent Financial Technology. It involved 213 financial professionals, 153 of them from external institutions, rating 579 real-world financial tasks. Hy4 scored 3.04, against 2.97 for DeepSeek-V4.1-Flash-1017 and 2.68 each for Kimi K3.1 Pro and GLM-5.3.
• Investment research. Tencent reports a 12.03 percentage-point gain over the preview on iRaB, an internal Investment Research Agent Benchmark run by the financial AI company Rabyte.
• Game development. Tencent reports gains over the preview of 19.1% on GameDev and 9.8% on GameLogic. OrcaRouter notes these are internal suites with no published item counts.
• WorkBuddy. On Tencent’s WorkBuddy evaluation, Hy4 scored an average of 83.18 with a 96.15% pass rate. Tencent says it ranked first in nine of twelve scenarios in the mainland Chinese market.
• Science. Tencent says it validated the model on TenureTrail, a set of 123 executable tasks across seven disciplines.

The News International summarises the wider self-benchmarking from Tencent’s technical blog. By that account, Hy4 beat the preview on 12 tasks and beat GLM-5.3 and DeepSeek-V4.1-Flash-1017 on all tasks, while results against Moonshot’s Kimi K3.1 Pro were mixed. The outlet cautions that these should be taken “with a grain of salt until third-party reviewers test them”. The only third-party data point in the sources is for the preview, which ranked 23rd on Code Arena’s WebDev leaderboard.

The hallucination figures do not match

Sources disagree on Tencent’s headline accuracy claim. Hallucination here means the model stating false information. Tencent’s announcement says the hallucination rate fell from 12.95% in the preview to 7.2%. OrcaRouter reports the figures as 15.76% to 5.45%. The two do agree that errors from unsupported assumptions fell from 23.1% to 5.6%. OrcaRouter adds that the methodology is unpublished.

Safety choices: filtered cyber training data

Tencent says it filtered training material on exploit development, reverse engineering and penetration testing out of Hy4’s training. It says it focused instead on defensive cybersecurity uses, to limit how much the model can help with attacks. The announcement ties the planned weight release to a focus on “cybersecurity and biological safety”. CXO Digitalpulse, citing Reuters, notes that Moonshot AI also restricted the cyber-offensive abilities of its latest model in October.

What it means for builders

For teams already testing the preview through an API, the full model brings vision and desktop control at an unchanged listed price. OrcaRouter calls computer use the more significant addition: “the first time the Hy line claims agency over a desktop rather than over a tool schema”. A tool schema is a fixed list of functions a model can call.

Teams that planned to self-host should wait. OrcaRouter’s assessment is that the main question is “whether you intended to run the weights yourself”. The preview is downloadable under Apache 2.0. The full model has no published weights, licence or model card yet.

Several limits apply:

• Tencent’s own stated gaps. Tencent acknowledges that Hy4 still lags the leading closed-source systems in research-level mathematics, long-horizon coding, model self-improvement and advanced visual question answering.
• No public coding scores. The release includes no standard public coding benchmarks. The preview’s published scores remain the best available stand-in for the family.
• Unmeasured efficiency claims. Without token counts, the shorter chain of thought is something to test on real workloads, not something to assume.
• Agent safeguards. SiliconFlow’s earlier advice for the preview still applies, and more so with computer use. Define tool permissions, timeouts, retry limits and points where a human must approve an action before giving an agent shell access.

What to watch next

• Weights and system cards. Whether they arrive “in the coming days” as promised, and under which licence.
• Specs. Whether the model card confirms the 770B/49B architecture and describes the vision encoder.
• Independent tests. Third-party results on public coding, vision and computer-use benchmarks.
• The hallucination figures. A methodology that explains the conflicting numbers.
• Hy4 Flash and Hy5. Tencent’s next releases are Hy4 Flash, a lighter version meant to cut task costs, followed by Hy5.

Spending is part of the backdrop. Tencent’s capital expenditure rose 176% year over year to 52.8 billion yuan (US$7.9 billion), according to The News International.