What can Tencent Hy4 read from one image?

Read a chart with a missing value

Select tencent/hy4 in an image-capable OpenRouter chat and attach orders.png. Copy the chart prompt from prompts.txt. It asks for the quantity and the exact labels above each bar. Friday’s label is deliberately covered. Save the first answer before opening answers.json.

1. Luke’s run returned orders, Wednesday 24, Thursday 36 and Friday unknown. The chart tells the reader one count is withheld, so this checks whether the model follows that instruction.

Rebuild the three objects

Start a fresh chat and attach only reference.png. Copy the scene prompt from prompts.txt. It supplies a six-unit horizontal span, a ground plane and three allowed shapes. Save the JSON, then compare the colours, order and dimensions with reference.json. The two Blender files let you inspect the original and estimated objects from the same camera.

1. Hy4 identified the blue cube, orange sphere and green cylinder. It estimated the cube at 2.20 units tall against a measured 1.35, about 63% too tall. The orange sphere measured 0.88 versus an estimate of 1.00 and the green cylinder measured 1.75 versus 2.40. Luke supplied the camera, materials, lighting and renderer.

Where the desktop scores come from

Tencent’s release chart reports 78.9 for Hy4, 75.2 for GPT-6.1 Sol and 81.8 for Claude Opus 5.5 on OSWorld 2.1. On Terminal-Bench 4.0 it reports 44.4 for Hy4 and 66.4 for Claude. Luke’s experiment used the hosted Hy4 on 7 November 2026. Four calls cost US$0.007632 in total, including two that ran out of output tokens.

1. Compare the estimated height to the original using (estimate − original) ÷ original × 100.

The takeaway

Hy4 read the visible labels accurately. The Blender comparison shows how a plausible reconstruction can still miss the measurements.