Tencent launches Hy4 with vision and computer use

The final version of Hy4 is here. Tencent announced it on November 4 with image input and the ability to act in terminals and graphical interfaces. It keeps the August preview’s standard API rates. The final model’s weights are still pending release, according to the announcement.

If the name sounds familiar, Hy4 preview was already available. Today’s release is its successor with vision. Check the version before downloading weights or comparing a benchmark screenshot.

What the Hy4 benchmarks show

Start with the task you care about. In Tencent’s table, Hy4 reaches 75.2 on APEX-Agents v1.1, the highest figure in that row. On Terminal-Bench 4.0 it scores 44.4, versus 66.4 for Opus 5.5 in the run marked as Tencent’s own. Switching from working agents to terminal coding changes the picture.

TENCENT APPENDIX · NOV 4, 2026

AgentsTerminalVisionDesktop

APEX-Agents v1.1

Pass@1 · higher is better

050100

1. Hy4 preview53.9
2. Hy475.2
3. GLM-5.364.3
4. DeepSeek-V4.1-Flash-101772.9
5. Kimi K3.1 Pro68.2
6. GPT-6.1 Sol61.7
7. Claude Opus 5.574.8

Terminal-Bench 4.0

Score · higher is better

050100

1. Hy4 preview26.3
2. Hy444.4
3. GLM-5.342.9
4. DeepSeek-V4.1-Flash-101740.0
5. Kimi K3.1 Pro66.8
6. GPT-6.1 Sol55.5
7. Claude Opus 5.566.4

MMMU-Pro

Vision · higher is better

050100

1. Hy4 previewNo result
2. Hy478.6
3. GLM-5.3No result
4. DeepSeek-V4.1-Flash-101776.7
5. Kimi K3.1 Pro83.4
6. GPT-6.1 Sol84.9
7. Claude Opus 5.588.2

OSWorld 2.1

Full · partial score, not completed tasks

050100

1. Hy4 previewNo result
2. Hy478.9
3. GLM-5.3No result
4. DeepSeek-V4.1-Flash-101768.4
5. Kimi K3.1 Pro79.6
6. GPT-6.1 Sol75.2
7. Claude Opus 5.581.8

Each model at its highest reasoning setting. When two figures are given, the chart uses the starred one, identified by Tencent as its own testing. Original cells are preserved below.

See original cells and methodology

Harnesses and configurations vary. Terminal-Bench uses Claude Code, up to 500 turns and 13 hours per attempt. OSWorld reports partial score. These tests do not form a single overall score.

APEX-Agents v1.1 · Pass@1 · higher is better

Model	Original cell
Hy4 preview	53.9
Hy4	75.2
GLM-5.3	64.3*
DeepSeek-V4.1-Flash-1017	72.1 / 72.9*
Kimi K3.1 Pro	67.2 / 68.2*
GPT-6.1 Sol	60.0 / 61.7*
Claude Opus 5.5	73.5 / 74.8*

Terminal-Bench 4.0 · Score · higher is better

Model	Original cell
Hy4 preview	26.3
Hy4	44.4
GLM-5.3	42.9*
DeepSeek-V4.1-Flash-1017	42.3 / 40.0*
Kimi K3.1 Pro	65.7 / 66.8*
GPT-6.1 Sol	55.9 / 55.5*
Claude Opus 5.5	66.4 / 66.4*

MMMU-Pro · Vision · higher is better

Model	Original cell
Hy4 preview	No result
Hy4	78.6
GLM-5.3	No result
DeepSeek-V4.1-Flash-1017	76.7*
Kimi K3.1 Pro	84.1 / 83.4*
GPT-6.1 Sol	84.9*
Claude Opus 5.5	88.2*

OSWorld 2.1 · Full · partial score, not completed tasks

Model	Original cell
Hy4 preview	No result
Hy4	78.9
GLM-5.3	No result
DeepSeek-V4.1-Flash-1017	68.4*
Kimi K3.1 Pro	79.6
GPT-6.1 Sol	75.2*
Claude Opus 5.5	81.8

Source and full appendix: Tencent Hy

The chart keeps a 0–100 scale. Each test has its own metric. OSWorld 2.1 reports partial score: 78.9 is that score, not a claim that Hy4 completes 78.9% of tasks. On the visual test MMMU-Pro, it scores 78.6.

The preview comparison helps explain the update: Terminal-Bench moves from 26.3 to 44.4; APEX-Agents from 53.9 to 75.2. Tencent notes that some preview figures were recalculated after changes to the evaluation setup. Read these as the values in this appendix.

Each model uses its highest available reasoning setting. Asterisks identify Tencent’s own testing, and configurations vary. Terminal-Bench allows up to 500 turns and 13 hours per attempt. HeyMark did not run these tests. Original cells and methodology are available below the chart.

API pricing and the rate observed on OpenRouter

These amounts are USD per million tokens. Standard rates come from Tencent; OpenRouter’s rates are those observed while reviewing the November 4 launch. Its listing does not give an end date for the price difference.

For stated billed totals of 10,000 input tokens and 2,000 output tokens, standard pricing comes to $0.013342: (10,000 × 0.834 + 2,000 × 2.501) / 1,000,000. At the observed OpenRouter rates it is $0.006671. This is an example using declared quantities; your requests can consume different tokens and tools.

OpenRouter identifies the version as tencent/hy4-20261104, with up to 1,048,576 context tokens and 65,536 output tokens. It accepts text and images and returns text. Reasoning effort can be high, low or none, with high as the default. Those limits describe this API offering; individual products can impose others.

What is pending before you can download final Hy4

Tencent says it plans to release the weights together with a system card. The announcement leaves the date and license unspecified. Check Tencent’s official Hugging Face profile to see which version you are downloading.

The 770B parameters, 49B active parameters and Apache 2.0 license belong to the Hy4 preview repository. That card documents August’s release. The final announcement does not disclose its architecture, so those specifications cannot automatically be assigned to final Hy4.

Tencent Hy’s original launch post on X, dated November 4. The full announcement and appendix explain the conditions behind the release.

A useful way to try the new visual input

For a content team, one practical test would be to supply a screenshot of a creative and its brief, then ask Hy4 to locate text, review hierarchy and suggest specific changes. Check its answers against the screenshot. This is our proposed test, not a capability independently validated by HeyMark or permission for a model to publish to your accounts.

If you already used the preview, compare both versions with the same material, instructions and budget. Record errors and the actual cost of completing the task. The added modality gives you a reason to try it; the benchmark table helps you decide what to test.