Notes on film, new tools and the decisions behind the work.

Tools / Analysis

Hy4 leaves preview with vision and computer use

Tencent's new release can work with images and software interfaces. Here is what changed from the text-only preview, how developers can access it, and why downloadable weights are a separate milestone.

A selection in a detailed aerial city view is enlarged as an architectural inset. Conceptual illustration of visual search, not Tencent demo footage or model output.
DU / SIGNALS · S/098Editorial illustration · AI-generated

Tencent released Hy4 on 4 November, adding image understanding and computer use to a model family whose preview accepted text. The release is available through Tencent Cloud and OpenRouter, and in Tencent products including CodeBuddy and WorkBuddy. OpenRouter lists text and image inputs, text output, and a 1,048,576-token context window. On 20 November, Tencent published the final model weights for self-hosting.

The consequential change is the link between seeing a task and acting on it. A model working on an interface can inspect its visual state as well as its code. Tencent says Hy4 combines command-line and graphical interaction, with targeted training for Blender, CAD and workflows that cross application boundaries.

Reading the screen changes the task

Tencent describes finer image inspection, chart interpretation and visual reasoning. Its launch page includes a task to locate an approximately two-metre rooftop component inside a 2.8-gigapixel aerial image. That is a concrete example of the intended work: finding a small relevant detail within a much larger visual source.

Computer use adds a separate execution layer. Accepting a screenshot through an API does not, by itself, give an application control of a desktop. Developers still need to connect the model to the tools that perform actions and return the next state. The useful test is whether the complete system can inspect a result, recognize an error and recover across several steps.

Sources: Tencent Hy [1] · OpenRouter [3]

X / DEMO

Tencent announces Hy4 with vision

Loading post from X…

Credit / Tencent Hy / @TencentHunyuanOpen original 

Tencent's official 4 November announcement links to the release page and demonstrations. It establishes the launch; it is not an independent performance test.

The target is work that takes more than one prompt

For game development, Tencent names Unreal Engine, Unity and Godot, alongside gameplay logic, state tracking and asset pipelines. It reports improvements of 19.1% on GameDev and 9.8% on GameLogic versus Hy4 preview. Those are gains on Tencent's evaluations, not measured reductions in a studio's production time.

Tencent also describes office-document work and research spanning several applications. These workflows require the agent to preserve earlier decisions while changing a document, interface or project. For a useful trial, give it an existing task with a checkable end state, ask for a revision, and inspect whether the change preserves the parts that should remain intact.

Sources: Tencent Hy [1]

An independent score, with a defined scope

Artificial Analysis gives Hy4 35 on its Intelligence Index, a composite covering several kinds of knowledge and reasoning work. It also lists text-and-image input and a one-million-token context window. The score is useful as an external comparison; it cannot establish that a particular Blender operation or a multi-day coding task will succeed.

Sources: Artificial Analysis [4] · Artificial Analysis [5]

The final weights are now downloadable

Tencent's launch price is $0.834 per million input tokens and $2.501 per million output tokens, with cached input at $0.042. On 6 November, OpenRouter still showed a 50% promotion: $0.417 input, $1.251 output and $0.021 cache read. These are provider-specific token rates; a long agent task can make several calls and generate substantial output.

Tencent published the final Hy4 weights on 20 November, along with a system card. The final model repository lists text, image and video inputs. Its language backbone has 770 billion parameters, with 49 billion active for each token. Those 49 billion describe the computation for a token. A self-hosted system still needs to store the full set of weights and keep them accessible during inference.

The earlier Hy4-preview repository remains a separate release. The final weights come under Tencent Hy License 1.0. Its terms include a separate licence requirement for qualifying large model-as-a-service businesses. Tencent’s public model card links to the full terms and the instructions for serving the model.

Sources: Tencent Hy [1] · OpenRouter [3] · Tencent on Hugging Face [6] · Tencent Hy / Hugging Face [7] · Tencent Hy / @TencentHunyuan [8] · Tencent [9]

X

Tencent releases the final Hy4 weights

Loading post from X…

Credit / Tencent Hy / @TencentHunyuanOpen original 

Tencent links to the weights, system card and licence in this announcement.

SOURCES & METHOD

The record behind the article

Based on Tencent's 4 November release article, its official announcement, OpenRouter's current model listing and Artificial Analysis. Product capabilities and internal evaluation results are attributed to Tencent. We have not run Hy4 or reproduced the benchmarks; the linked launch material is the developer's presentation.

  1. Introducing Hy4 Tencent Hy · 2026-11-04
  2. Hy4 is here, new with vision Tencent Hy / @TencentHunyuan · 2026-11-04
  3. Tencent: Hy4 — API pricing and providers OpenRouter
  4. Tencent Hy4 — Intelligence, Performance & Price Analysis Artificial Analysis
  5. Intelligence Benchmarking Methodology Artificial Analysis
  6. Hy4-preview — earlier text-generation model Tencent on Hugging Face
  7. Hy4 final weights and model card Tencent Hy / Hugging Face · 2026-11-20
  8. Tencent announces the downloadable Hy4 weights Tencent Hy / @TencentHunyuan · 2026-11-20
  9. Tencent Hy License, version 1.0 Tencent