Tencent Hy4: the 40-second productivity teaser

Tencent Hy4's November 3 teaser, upcoming general release and Flash: visual understanding, long tasks, and how these differ from the open-weight Hy4 Preview.

Updated: 2026-11-03

Tencent is warming up the Hy4 general release. At 20:15 Beijing time on November 3, @TencentHunyuan posted a 40-second film with the caption “4 Real Productivity.” The more concrete release signal comes from Tencent's October 29 announcement, which names both the upcoming Hy4 general release and Hy4 Flash. Official teaser ↗

The distinction matters: Hy4 Preview already exists. This new promotion should be read alongside the forthcoming releases, without transferring Preview's size, context window or license to them. This article reports verified sources and TOKRACE's interpretation; it contains no new model tests.

The official 40-second teaser

Official film and poster: Tencent Hy · 1080p · 40 seconds. Download MP4 · Official post ↗

The film moves through natural patterns and scientific imagery before reaching a desk chart, a microscope, a constellation and a small electronics board. Near the end, a notebook connects reasoning with action; “To See,” “To Work” and “To Research” flash across the screen. It closes on Tencent Hy 4.

Our reading is a shift toward usable work: seeing a situation, acting on it and carrying an investigation forward. The montage gives that positioning a visual language. It does not show a complete software session or establish scientific accuracy, financial prediction, native video generation, or an autonomous task success rate.

What Tencent has actually announced

In its October 29 corporate announcement, Tencent says the upcoming Hy4 general release focuses on long-horizon and real productivity tasks, more efficient reasoning, and visual understanding. It also describes Hy4 Flash as a cost-efficient model retaining visual and agent capabilities. These are Tencent's announced directions, not TOKRACE measurements. Tencent announcement ↗

01 / SEE

Read the situation

Visual input could let an agent reason about screenshots, diagrams and interfaces. Understanding an image still needs to be distinguished from generating one.

02 / WORK

Keep a task moving

Long tasks require a model to retain constraints, use tools and continue after errors. A polished final picture reveals only part of that process.

03 / EFFICIENCY

Reach a result sooner

Reasoning efficiency matters when checking and revising dominate a task. More output tokens alone do not establish more useful work.

The three points above are our interpretation of why those directions matter. Together they suggest evaluating an agent by the work it completes across inputs, tools and revisions, with time and cost attached to the whole task.

Hy4 Preview, Hy4 and Flash are different entries

Tencent announced and open-sourced Hy4 Preview on August 28. Its repository documents the model and weights. The table separates that existing release from the two upcoming entries. Preview release announcement ↗

Version	Status in the checked sources	What can be established
Hy4 Preview	Already released; open weights	770B backbone / 49B active; 1M context; Apache 2.0. These specifications belong to Preview.
Hy4 general release	Upcoming in the October 29 announcement	Long-horizon work, reasoning efficiency and visual understanding are named directions; exact release specifications remain to be announced.
Hy4 Flash	In preparation in the same announcement	Positioned for cost efficiency with visual understanding and agent capabilities; no price established by these sources.

Preview specifications: Tencent-Hunyuan/Hy4-preview. The 770B figure excludes the separate MTP layer; it is the backbone figure used by the repository.

What would make this useful in practice?

Visual understanding becomes useful when it changes the next action. For example, a model reading a chart must retain units and labels when it uses the data; a model inspecting an interface must connect what it sees to the right controls. These are task requirements, not examples demonstrated by this teaser.

Long-horizon work adds another constraint: a change late in the process must still respect the original request. That is why a useful evaluation preserves the editable project and checks what happens after a requested revision. A teaser can introduce the ambition; a finished, inspectable artifact is where that ambition becomes testable.

TOKRACE's existing Kimi creative evaluation and LongCat 2.5 work samples show how we examine scenes, videos and source projects. Their results apply to those models and runs; this Hy4 article is release coverage.

Tencent Hy4 release questions

Has Tencent Hy4 launched?

Hy4 Preview is already available. The official materials checked on November 3 describe the Hy4 general release and Hy4 Flash as upcoming; they do not give a specific release date.

Are 770B parameters and 1M context confirmed for the general release?

These are Hy4 Preview backbone specifications. The sources checked here do not establish those specifications for the forthcoming general release or Flash.

Where can I try Hy4 now?

Tencent Cloud and Tencent AI Studio provide official model entry points, while Preview weights are available on Hugging Face. Check the exact version shown in each product; a Preview entry is not confirmation of access to the general release.

Tencent Cloud Hunyuan ↗ · Tencent AI Studio ↗ · Hy4 Preview / Hugging Face ↗

Sources and attribution

Checked November 3, 2026 (Asia/Shanghai). Official claims and TOKRACE interpretation are identified in the relevant sections. The official MP4 and poster are hosted on TOKRACE with source URLs and file hashes retained; the social card is TOKRACE's own typographic illustration.

Post text, timestamp and duration were verified through X's public syndication metadata. Source and media manifest · More model news and reports