Tencent Hy4 review: vision, Blender, SVG and knowledge videos

Tencent Hy4 on OpenRouter: five image-input tests and the standard creative suite, with original inputs, answers, rendered works, source projects and attributed official benchmarks.

Updated: 2026-11-04

Meet Tencent Hy4

Tencent Hunyuan logo

Tencent Hy4 · 腾讯混元

Tencent released the final Hy4 on November 4, 2026. The official release adds visual understanding and computer use, with an emphasis on fine image details, complex charts and longer tool-based tasks. Official release

The OpenRouter route used here is tencent/hy4. Its November 4 catalog accepts text and images, returns text, and lists a 1,048,576-token context and a 65,536-token maximum output. The official post lists weights and a system card as planned releases. OpenRouter

The OpenRouter quote captured on November 4 was $0.417 per million input tokens, $1.2505 for output and $0.021 for cached input, with a 50% discount whose end date was not disclosed. Measured usage and API-reported costs are recorded separately below.

Read the Hy4 launch analysis

Results overview

All eleven creative tasks delivered artifacts. SVG constraints, shelf dimensions and sampled gear motion hold up; manufacturing fits and some scientific video diagrams remain incorrect. The separate image-input set passes all 25 checks, without a matched Preview comparison or a measured improvement claim.

This page contains 11 creative task records and 5 separate image-input tests. The creative works were made through Claude Code; the image-input tests used independent OpenRouter requests without tools.

Image understanding: original inputs and answers

25/25 · 5 tasks, 6 synthetic PNGs

These are first responses with no tools or external OCR, scored by a coordinating AI against frozen checks. The occlusion task explicitly says the covered information is unrecoverable; its result establishes compliance with that boundary under this instruction.

Chinese sheet and unit-aware arithmetic 5/5

Correctly reads Chinese labels and prices, applies both row filters, and converts packages into units: 324 items and CNY 1,788.00.

Original task, answer and request details

Original prompt (Chinese)

请仅依据附件图像,读取这张待出库复核单。1)抄出 K-18 的品名和单价/件;2)找出同时满足“未复核”和“优先=是”的所有行,给出货号、实际件数和金额;3)汇总这些行的实际总件数、总金额。请写出包装数和每包件数的换算过程,金额保留两位小数。不要把表中的包装数直接当件数。 请按问题编号简洁作答,保留可核对的计算或路径。

Original model answer (Chinese)

1)K-18:品名「透明收纳盒」;单价/件 15.80 元/件。

2)筛选“复核=未复核”且“优先=是”,得到 K-17、C-09:

货号包装数 × 每包件数 → 实际件数金额
K-176 × 4 = 24 件24 × 29.50 = 708.00 元
C-0915 × 20 = 300 件300 × 3.60 = 1,080.00 元

3)汇总:

  • 实际总件数:(6×4) + (15×20) = 24 + 300 = 324 件
  • 总金额:(6×4×29.50) + (15×20×3.60) = 708.00 + 1,080.00 = 1,788.00 元

Download the original answer

Requested / returned model: tencent/hy4 / tencent/hy4 · Tencent

1,864 / 2,239 input / output tokens, including reasoning · 24.69 seconds

inputs/warehouse-sheet.png
SHA-256: 48035df4f424714042c09a1491fc74d511c849346e2989aff37f8facf34360f7

Dual-axis chart and weighted reasoning 5/5

Reads all six months on both axes correctly; identifies the returned-volume peak and computes 28.57% growth and a 3.69% weighted overall return rate.

Original task, answer and request details

Original prompt (Chinese)

请仅依据附件图表回答:1)6月订单量和退货率各是多少,分别用什么单位?2)估算退货量最高的是哪个月,是多少千单?3)5月到6月订单量环比增长百分之几?4)这6个月合计的退货率是多少?请解释为何不能把6个百分数直接求平均,列出所读到的每月数据和计算过程,百分比四舍五入到小数点后两位。 请按问题编号简洁作答,保留可核对的计算或路径。

Original model answer (Chinese)

1)6月订单量为 180千单(蓝柱/左轴,单位“千单”,即18万单);退货率为 2.00%(红线/右轴,单位“%”/百分比)。

2)按图注“估算退货量=当月订单量×当月退货率”,逐月读数及计算如下(订单量、退货量单位:千单):

月份订单量退货率估算退货量
1月805.00%80×5.00%=4.00
2月1204.00%120×4.00%=4.80
3月1006.00%100×6.00%=6.00
4月1603.00%160×3.00%=4.80
5月1404.00%140×4.00%=5.60
6月1802.00%180×2.00%=3.60

最高是 3月,6.00千单。

3)5月到6月订单量环比增长: [ \frac{180-140}{140}\times100%=28.5714%\approx\boxed{28.57%} ]

4)合计订单量: [ 80+120+100+160+140+180=780.00\text{千单} ]

合计估算退货量: [ 4.00+4.80+6.00+4.80+5.60+3.60=28.80\text{千单} ]

按图注以订单量加权,整体退货率: [ \frac{28.80}{780.00}\times100% =\frac{\sum(\text{月订单量}\times\text{月退货率})}{\sum\text{月订单量}} \approx\boxed{3.69%} ]

不能直接平均,因为每个月退货率的分母——订单量——不同;直接平均相当于假定各月订单量相同。直接平均为: [ \frac{5.00%+4.00%+6.00%+3.00%+4.00%+2.00%}{6}=4.00% ] 它不等于按订单量加权后的整体退货率 3.69%。

Download the original answer

Requested / returned model: tencent/hy4 / tencent/hy4 · Tencent

1,861 / 4,240 input / output tokens, including reasoning · 46.18 seconds

inputs/orders-returns.png
SHA-256: 87ae914e35737630c7c64a83ea48cc5492fe5587005a6d17162be31f56cece0e

Floorplan counting and route planning 5/5

Correctly counts six sensors and applies the spatial filter. Its 17-move coordinate route passes adjacency and obstacle checks and matches the shortest path.

Original task, answer and request details

Original prompt (Chinese)

请看附件仓库巡检平面图:1)蓝色传感器共有几个?2)在E列以东且出口以北的传感器有几个,各在哪一格?3)从入口A1到出口G6,在图示规则下最短需要移动几步?给出一条最短路线的逐格坐标,并解释必须绕过的主要障碍。传感器所在白格可以通行;不许穿过深色格或斜走。 请按问题编号简洁作答,保留可核对的计算或路径。

Original model answer (Chinese)

1)6个:B1、D1、F1、B4、D5、G5。

2)2个:F1、G5。E列以东为F/G列,出口G6以北为第1–5行。

3)最短17步(18个坐标、17次上下左右移动):

A1→A2→A3→A4→B4→B5→C5→D5→D4→D3→D2→E2→F2→F3→G3→G4→G5→G6

步数核验:A1→D5为7步,D5→D2为3步,D2→G3为4步,G3→G6为3步,共
[ 7+3+4+3=17. ]

主要绕障:C1、B2/C2、B3/C3、C4形成左上暗格墙,须下到B4/B5、经C5绕至D5;E3–E6及F4/F5又形成中央墙,D侧只能在第1、2行向东,故沿D列北上至D2,经E2、F2、F3转入G列。F6受E6/F5封成死端,因此须从G5进入出口;B4、D5、G5虽为传感器格,但可通行。四邻接BFS中两个关口距离分别为 (d(E2)=11,d(E1)=12),而E2到G6的曼哈顿下界为6步且上述路线达到,故17步确为最短。

Download the original answer

Requested / returned model: tencent/hy4 / tencent/hy4 · Tencent

1,853 / 13,219 input / output tokens, including reasoning · 152.85 seconds

inputs/inspection-floorplan.png
SHA-256: 2bb2550aaecfe63d947c795bb6b4d9ac591e2c2fe602a6852a2467d8fd1f9ccb

Multi-image visual change localization 5/5

Finds all five changes in the display, status, button position, box count, and valve orientation without inventing extras or a direction of rotation.

Original task, answer and request details

Original prompt (Chinese)

第一张是工位调整前,第二张是调整后。请比较两张图像,逐条列出所有明确可见的变化。每条说明所在区域、具体物件或文字、调整前和调整后的状态;把同一物件的文字及颜色改变算作一处变化。不要推断画面没有显示的故障原因或生产结果。 请按问题编号简洁作答,保留可核对的计算或路径。

Original model answer (Chinese)

  1. 控制台/“本周完成”显示值:128 → 132,132−128=4。
  2. 控制台/状态按钮(文字与颜色合并计一处):浅绿色“运行” → 橙黄色“待检”。
  3. 控制台/台面红色圆形按钮:由台面左端、左支腿内侧上方 → 台面右端、右支腿内侧上方(左侧 → 右侧)。
  4. 周转料架/下层料盒:调整前0个,调整后左侧有1个蓝色带白色短线的料盒;上层为3个不变,总数 3 → 4,4−3=1。
  5. 气路阀门/橙色长条阀柄:横向 → 纵向,方向相差90°;中央深棕圆点仍在圆心,顺/逆时针旋转方向无法由两图判定。

Download the original answer

Requested / returned model: tencent/hy4 / tencent/hy4 · Tencent

3,573 / 5,438 input / output tokens, including reasoning · 66.77 seconds

inputs/workstation-before.png
SHA-256: 61d398d1c8a9d99db2e187fdea14b8068978d3f4255fc3150fd7ea6bdb6a474b
inputs/workstation-after.png
SHA-256: 4640c8d2562589c9af181cadb6bba4db55ee7876cc4f73e38348fc6084f88128

Occlusion and uncertainty boundaries 5/5

Reads the visible fields and tracking fragments without inventing hidden characters. The image explicitly says the covered content is unrecoverable, so this checks boundary-following under that instruction.

Original task, answer and request details

Original prompt (Chinese)

请仅根据附件标签读取:收件城市、收件区域、分拣区码、件序,以及完整运单号。对每一项标明能否从图像确定。若某项被遮挡,分别抄出可见片段并说明哪些内容无法判断,不要猜测或按格式补齐。 请按问题编号简洁作答,保留可核对的计算或路径。

Original model answer (Chinese)

  1. 收件城市:能确定 → 南京市(标签第 1 行「收件城市:南京市」)。
  2. 收件区域:能确定 → 鼓楼区(第 2 行「收件区域:鼓楼区」)。
  3. 分拣区码:能确定 → A-204(第 3 行左侧「分拣区码:A-204」)。
  4. 件序:能确定 → 2/3(第 3 行右侧「件序:2/3」)。
  5. 完整运单号:不能确定。第 4 行「运单号:」后的可见片段仅为:左段「XH 5271」、右段「83」;两段之间为深色遮挡矩形,无可见像素。依标签说明「遮挡区域没有可恢复的信息;只依据可见像素回答」,无法判断遮挡部分的字符个数与内容,也无法判断整号总长度、左右片段之间是否另有字符或分隔符,故不拼接、不按格式补齐,完整运单号不可读。

核对路径:按标签正文自上而下逐行读取;仅第 4 行存在遮挡,其余四行均无遮挡。

Download the original answer

Requested / returned model: tencent/hy4 / tencent/hy4 · Tencent

1,492 / 1,109 input / output tokens, including reasoning · 12.21 seconds

inputs/partially-occluded-label.png
SHA-256: 541fe3f948d1127c88b3aff5ca16f0caa928f5235a46adce18f147b57321ff42

3D modelling

0.8 m three-bay shelf, main view

Parametric bookshelf and parameter revision

Both sizes, 18 mm panels and equal clear bays are correct, with readable light-oak materials and dado details. The comparison header has missing arrow glyphs.

Source project

More views and test details
1 m four-bay shelf, main view

1 m four-bay shelf, main view

Version one, rear view

Version one, rear view

Version two, rear view

Version two, rear view

Version one, structural view

Version one, structural view

Version two, structural view

Version two, structural view

Version one, dado detail

Version one, dado detail

Version two, dado detail

Version two, dado detail

Parameter comparison (missing header arrow glyphs)

Parameter comparison (missing header arrow glyphs)

37m 32s · Requests / complete responses: 70 / 69 · 5,628,469 / 104,828 input / output tokens

  • Actual meshes match both requested sizes, 18 mm panels and equal clear bays. One parameter script preserves editable boards, dado joints and dowels, with complete front, rear and structural views.
  • Light oak and consistent camera placement make the width and bay-count changes easy to compare. Arrow glyphs are missing in the optional comparison header; the six required views remain readable.
  • Parameter changes require rerunning the external Python builder. Rebuilding elsewhere also requires adjusting the Blender launcher path and providing Pillow and fonts.

Itemized verification

Measured dimensions and equal bays · pass
Outer sizes are 0.8/1.0 × 0.3 × 1.2 m, with 18 mm panels and clear bay heights of 376 / 277.5 mm.
Native projects and public copies · pass
Both originals reopened with auto-execution disabled. Public copies only remove embedded texts; object, mesh, material, camera and animation fingerprints match.
Six required views · pass
Both main/rear pairs are 2400×1350, with 1920×1080 structural views; all were visually inspected.
Comparison typography · partial
Two arrow glyphs in the optional comparison header appear as boxes.

Task source

Uses creative-suite-v4. The brief here is a summary; the hash identifies the actual archived creation prompt and environment.

persistent-author-session · native default · model_finished · Requests / complete responses: 70 / 69 Run record

One persistent Hy4 session authored the source, renders and public project exports; missing usage is listed in the run record.

Source brief and prompt
题面概要:同一脚本生成0.8×0.3×1.2米三格书架,再只改宽度为1米、分格为四格;板厚18毫米,每版交主、背、结构视图及原生工程。
提示词概要:使用同一参数脚本完成两版真实可编辑书架,核对尺寸、等高分格与接合,并实际渲染主视图、背面及结构近景。

SHA-256: 8c9b4f2b770b6e4a7b00add210fb2f1d5a36e45688e2f318cc4e82632085d36a

Assembled reducer

24/18/60-tooth planetary reducer and axial explosion

Tooth counts, involute geometry and the 3.5 transmission ratio are correct, with four views, mechanism animation and three LODs. Manufacturing fits and some callouts remain imperfect.

Source project

More views and test details
Axial exploded view

Axial exploded view

Gear meshing from above

Gear meshing from above

Involute tooth detail

Involute tooth detail

Music and sound effects, no narration.

Download MP4

Six-second mechanism motion and reversible turntable

104m 0s · Requests / complete responses: 162 / 161 · 13,919,531 / 231,020 input / output tokens

  • Actual meshes confirm 24/18/60 teeth, 20-degree involutes, 20 mm faces and 42 mm planet placement. Sampled motion satisfies the 3.5 ratio, and the four views plus cutaway animation explain the assembly.
  • The front plate obscures some gears in the hero view, and some callouts land on pins or flat faces. The top view clarifies meshing; the static GLB LODs contain 68,684, 41,420 and 26,476 triangles.
  • Manufacturing fits remain imperfect: the approximately 12 mm sun bore and 11.96 mm input shaft form clearance rather than H7/r6 interference. UVs have usable area, but full non-overlap was not independently verified.

Itemized verification

Gear geometry, dimensions and motion · pass
Actual vertices give 24/18/60 teeth and 20-degree involutes. Seven sampled poses have no planar tooth penetration, with a 3.5 sun/carrier angular ratio.
Final media and native project · pass
Four 1920×1080 images were viewed. The six-second 1280×720, 30 fps film fully decodes with audio and visible mechanism motion. The native project reopens with auto-execution disabled and has no embedded text.
Manufacturing fits · fail
The approximately 12 mm bore and 11.96 mm shaft form clearance rather than the labelled H7/r6 interference fit.
UV and LOD delivery · partial
Three GLBs provide UVs/normals and 68,684 / 41,420 / 26,476 triangles. UV non-overlap was not fully verified; GLBs contain no baked textures or animations.
Presentation and rendering · partial
Finals use EEVEE. The exploded view is a distant axial view; the hero front plate hides some gears and several callouts are imprecise.

Task source

Uses creative-suite-v4 creation conditions. This is a brief; the hash identifies the actual archived prompt and environment.

persistent-author-session · native default · model_finished · Requests / complete responses: 162 / 161 Run record

One persistent Hy4 session authored source and media; missing usage is identified in the sanitized record.

Source brief and prompt
题面概要:按附件制作24/18/60齿、模数2毫米行星减速器游戏资产,保持140毫米外径、3.5速比、UV与LOD;交装配、轴向爆炸、俯视、齿部图及六秒视频。
提示词概要:用真实渐开线几何和毫米尺度制作可编辑行星减速器,完成四视图、机构运动及可导出的游戏资产。

SHA-256: 4f2f0c15121c1c40e85e429628b4dbb98c4ca7ef6e6cbe9ce948cbcd03c4f2a7

Living and dining room

12×9m two-bedroom interior, rooms and floor plan

Layout, materials and mixed lighting are readable, with a complete floor plan and turntable. Master-bedroom framing is tight and furniture is simplified; output follows the run's shared HD conditions.

Source project

More views and test details
Master bedroom close-up (tight framing)

Master bedroom close-up (tight framing)

Second bedroom and study

Second bedroom and study

Final raw orthographic floor plan

Final raw orthographic floor plan

Labelled floor plan (derived from an earlier low-resolution preview)

Labelled floor plan (derived from an earlier low-resolution preview)

Clay structure and door/window openings

Clay structure and door/window openings

L-shaped kitchen

L-shaped kitchen

Alternate living/dining view

Alternate living/dining view

Master suite detail

Master suite detail

Music and sound effects, no narration.

Download MP4

Six-second roofless turntable, music without narration

120m 27s · Requests / complete responses: 141 / 139 · 11,911,141 / 257,801 input / output tokens

  • The orthographic plan and turntable clearly show living/dining, two bedrooms, kitchen, wardrobe and both bathrooms. Wood, tile, cabinetry and warm mixed lighting are distinguishable, with room-based native collections.
  • Both master-bedroom views are tightly framed and do not show the whole bed or room. Furniture, textiles and plants use simplified forms, and the windows lack exterior context.
  • The run meets the shared HD minimums. The original attachment's 4K and all-16-bit target was not reproduced under these shared production conditions. The labelled plan enlarges an earlier low-resolution preview; the final orthographic render is listed separately.
  • The native project, relative texture assets and author scripts are supplied. No complete rebuild on another machine or exhaustive collision check was performed; audio received technical checks without a listening review.

Itemized verification

Native project and textures · pass
The final 0494a10… project reopens with auto-execution disabled: 731 objects, 18 textures, 22 cameras, 39 lights, and no embedded text.
Required views and video · pass
The three interior views, raw orthographic plan, clay view and extra views were inspected. The six-second 1280×720/30 fps film fully decodes with audio and visible turntable motion.
Resolution and bit-depth conditions · partial
Meets the run's main-view minimum of 1920×1080 and other-view minimum of 1600×900. The attachment's 4K/all-16-bit target was not reproduced; most colour views are 16-bit, while both plans are 8-bit.
Framing and floor-plan detail · partial
Master-bedroom framing is too tight. The labelled plan enlarges an early low-resolution image; its 2400×1600 file dimensions do not establish equivalent rendered detail.
Exhaustive collisions and cross-machine rebuild · not-tested
Furniture collisions, every door width and a complete rebuild on another machine were not exhaustively verified.

Task source

Uses this creative-suite-v4 run's shared minimum resolution and production conditions. This is a brief; the hash identifies the actual archived prompt and environment.

persistent-author-session · native default · model_finished · Requests / complete responses: 141 / 139 Run record

One persistent Hy4 session authored the project, frames and audio scripts; the coordinator executed its existing encoder. Missing call usage is identified separately.

Source brief and prompt
题面概要:12×9米现代简约两居室,净高2.8米,按附件配置客餐厅、厨房、东北主卧套间、东南次卧书房及公卫;交室内、顶视、灰模和六秒视频。
提示词概要:制作可编辑两居室,保留真实门窗洞口、房间分组与PBR材质,完成室内视角、正交平面、灰模及六秒转台。

SHA-256: a7390deaf4be738147e248f4ab6f878d3cc4f63b3add06f7b59c88978190fcc5

Knowledge videos

Why ice floats: an error in the buoyancy diagram

The floating scene overstates displaced-water weight and volume. Molecule and density diagrams are clear, and the 83.51-second film and source files are complete.

Source project

Test details

57m 16s · Requests / complete responses: 86 / 85 · 7,045,916 / 125,673 input / output tokens

  • The molecule → volume → density explanation is clear, with readable diagrams and Chinese captions.
  • The buoyancy scene confuses equal-volume density comparison with floating equilibrium and depicts the displaced volume incorrectly.
  • Chinese narration, subtitles, music and source files are delivered. Technical and timing checks are complete; audio has not been reviewed by listening.
  • Author source retains local absolute Playwright/Chromium/TTS paths that require adjustment on another machine. Packaged raw narration lets users skip TTS, which refuses to overwrite existing output.

Itemized verification

Complete film and format · pass
Full audio/video decode passes: 1920×1080, 30fps, 2,505 frames, 83.51-second container, 48kHz stereo AAC.
Molecular arrangement and mass conservation · pass
Both phases show hydrogen bonds, with flexible liquid bonds and an open ice lattice. The same 24 molecules are retained and mass loss is explicitly rejected.
Approximate 9% expansion and 4°C scope · pass
About 9% refers to volume expansion, and roughly 4°C to freshwater maximum density. Diagrams are marked schematic and the saltwater caveat is retained.
Floating equilibrium and displaced volume · fail
Around 50–66 seconds, the film says displaced water is heavier than the ice. At floating equilibrium their weights are equal. Its yellow displaced-volume box reaches the cup bottom instead of matching the submerged ice volume.
Captions and timeline · pass
All 17 SRT cues have valid times, from 0.55 to 82.58 seconds. Text matches the narration, with scenes sharing the sample-based edited-audio timeline; pronunciation has not been verified by listening.
Audio track and mix · partial
Raw Kokoro text and audio hashes match their receipt. Trimming sentence-end silence reduces 93.1 seconds to about 83.51 without tempo acceleration. Voice RMS is about −21.26dBFS and BGM −32.52dBFS; no listening review was performed.

Task source

Uses the frozen November 4 topic and narration guide with the shared Kokoro voice tool. This brief is a summary; the hash identifies the original session prompt.

NASA Glenn: buoyancy and floating-equilibrium check

Transport Canada: buoyancy and floating-equilibrium check

persistent-author-session · native default · model_finished · Requests / complete responses: 86 / 85 Run record

One persistent Hy4 session authored the work; run metadata and missing usage are recorded separately.

Source brief and prompt
题面概要:依据当天冻结资料讲清普通冰为何浮于水面,中文口播、字幕与机制动画,60–90秒、1920×1080,先测实际语音时长。
提示词概要:由Hy4按冻结事实和口播指南写讲稿/读音稿/分镜,生成中文配音、连续机制动画和同步字幕,并交成片与源码。

SHA-256: 8a1fbeef2c005cc9963617131575ba59cc442eedb80bf80c816d0c6a99a518d0

QR error correction: clear mechanism, mismatched small-diagram count

A small 50-codeword diagram contains 40 cells, with a minor legend collision. The error-correction explanation and recovery-rate limits are clear; the 86.375-second film and source are complete.

Source project

Test details

30m 20s · Requests / complete responses: 83 / 83 · 6,986,137 / 86,434 input / output tokens

  • A small diagram around 52 seconds labels 40 cells as 50 codewords; the legend and footnote need tidying.
  • The film clearly separates schematic mechanisms from actual scanning; key percentages and the 200/50 example are correct.
  • The film, Chinese narration, 13 captions and source files are delivered. Full decode and timeline checks pass; audio has not been reviewed by listening.
  • Source depends on macOS fonts and a local Kokoro path. Another machine needs Pillow/NumPy/SciPy, FFmpeg and adjusted font/tool paths; packaged raw narration and mix can be reused.

Itemized verification

Complete film and format · pass
Full audio/video decode passes: 1920×1080, 30fps, 2,591 frames, 86.375 seconds, with 48kHz stereo AAC.
Error correction and schematic scope · pass
Explains 8-bit codewords, Reed–Solomon parity relations and recovery from redundancy; explicitly labelled schematic rather than a real scanning experiment.
200/50 example and four correction levels · pass
The main 20×10 grid has 50 recovered positions. Q at 25% and L/M/Q/H at about 7/15/25/30% are correct, distinguished from arbitrary image area and online retrieval.
Small-diagram count and legend · partial
At 52 seconds, a 50-codeword card draws 8×5=40 cells. Around 62.8 seconds, the explanatory footnote touches legend outlines.
Captions and measured-audio timeline · pass
All 13 subtitle cues have valid times, from 0.04 to 85.415 seconds. Scenes use actual TTS segment samples; caption text matches the narration.
Audio and mix · partial
The 85.475-second Kokoro track matches text/audio receipt hashes. It is resampled to 48kHz and padded by 0.9 seconds without tempo change. Original synthetic BGM/SFX are below voice; no listening review.

Task source

Uses the frozen November 4 topic and narration guide with the shared Kokoro tool. This brief is a summary; the hash identifies the original session prompt.

DENSO WAVE: QR error correction and codeword ratios

persistent-author-session · native default · model_finished · Requests / complete responses: 83 / 83 Run record

One persistent Hy4 session authored the work; actual API usage is preserved in a separate sanitized run record.

Source brief and prompt
题面概要:依据当天冻结的DENSO WAVE资料解释二维码纠错冗余与损坏边界,制作60–90秒、1920×1080中文口播、同步字幕及机制动画。
提示词概要:由Hy4按冻结事实与口播指南写讲稿、读音稿、分镜,生成中文旁白、同步字幕与纠错机制动画,并交付成片与可复建源码。

SHA-256: 031d0f7ae37094aecf7073787e63c274050aa8751770245452f6edc427357313

Webb insulation: five membranes miscounted as five gaps

The closing miscounts four inter-layer gaps as five, and overstates the thick-plate comparison. The complete 87-second film has clear hot/cold orientation, sideways-radiation diagrams and a material cross-section.

Source project

Test details

31m 5s · Requests / complete responses: 63 / 62 · 5,020,566 / 100,166 input / output tokens

  • Five membranes are described as five gaps; a relative thick-plate comparison becomes an absolute claim.
  • Radiation through vacuum, sideways heat escape, materials and the layer-count tradeoff are clearly visualized.
  • The 87-second film, 26 captions, voice and source are complete. Sub-cues use character ratios within sentences; no listening review was performed.
  • Source retains local Node/Playwright/Chromium/TTS paths and macOS fonts that need adjustment elsewhere. Raw voice and mix are included; the controller and temporary frames are excluded, and render.mjs can be called directly.

Itemized verification

Complete film and format · pass
Full audio/video decode passes: 1920×1080, 30fps, 2,610 frames, 87.00 seconds and 48kHz stereo AAC.
Vacuum, radiation and direction · pass
Separates contact conduction from heat radiation through vacuum. Sun/hot side is above, telescope/cold side below, with heat escaping sideways between layers.
Membrane and inter-layer-gap count · fail
Closing narration, captions and graphics say five gaps. Five membranes have four inter-layer gaps; the animation places a fifth counting ring below the lowest membrane.
Materials and layer-count tradeoff · pass
Kapton, aluminum on both sides and treated silicon on the sunward face of the two hottest layers are readable. The 2–7-layer study and cooling/redundancy/weight/complexity tradeoff are explained without invented temperature differences.
Wording of the thick-plate comparison · partial
Saying a thick plate has no exit for heat and thickness is no defense overstates NASA’s relative comparison of separated membranes and a solid thick plate.
Caption timing and synchronization scope · partial
All 26 SRT cues have valid times, ending at 86 seconds. Sentences use real Kokoro sample counts; smaller cues are positioned by character ratios within each sentence, not word timestamps or listening review.
Audio and mix · partial
The original 86-second Kokoro voice matches its text/audio receipt hashes. It is not accelerated; the mix is padded to 87 seconds with quieter BGM/SFX. No listening review was performed.

Task source

Uses the frozen November 4 topic and narration guide with the shared Kokoro tool. This brief is a summary; the hash identifies the original session prompt.

NASA: Webb sunshield, heat transfer and layer tradeoff

persistent-author-session · native default · model_finished · Requests / complete responses: 63 / 62 Run record

One persistent Hy4 session authored the work; observed usage is stored separately, and missing usage is not counted as zero.

Source brief and prompt
题面概要:依据当天冻结的NASA资料,解释韦布为何使用五层分离薄膜而非一整块厚板,制作60–90秒、1920×1080中文口播、字幕及机制动画。
提示词概要:由Hy4依据冻结的来源事实与口播指南写讲稿、读音稿、分镜,解释真空/辐射/材料与层数权衡,交付中文配音、同步字幕、成片和可复建源码。

SHA-256: 1b5386eab431faa721b0fd148bb57429305ec1ccd41cf290f69a370b8a6739e1

Noise-cancelling headphones: an error in wave addition

The addition diagram gives a nonzero residual from perfectly opposite waves, while the opening links pausing music to a drop in external hum. The loop and seal limitations are clear, and the 88.167-second film and source are complete.

Source project

Test details

31m 50s · Requests / complete responses: 50 / 49 · 3,101,897 / 82,693 input / output tokens

  • The purple residual is not the sum of the equal-amplitude opposite waves; the opening also suggests pausing music lowers external hum.
  • The loop, physical seal, local cancellation and limits for steady hum and speech are clearly explained.
  • The 88.167-second film, 15 captions, voice and source are complete. Full decode and subtitle-stream checks pass; audio has not been reviewed by listening.
  • Source retains local Playwright/Chromium/TTS paths and macOS fonts that need adjustment elsewhere. Raw voice, final mix and audio-generation code are included; inputs, hidden controller state, logs and temporary frames are excluded.

Itemized verification

Complete film and format · pass
Full audio/video decode passes: 1920×1080, 30fps, 2,645 frames, 88.166667-second video/container and 88.15-second 48kHz mono AAC.
Mathematical addition of opposite-phase waves · fail
Around 47–54 seconds, equal-frequency, equal-amplitude sinusoids separated by pi are shown producing a residual. They cancel exactly; the purple residual is drawn independently rather than computed as their sum, contradicting the plus/equality symbols.
Pausing music and external noise · partial
The opening pairs music-paused graphics with falling amplitude labelled external steady hum, suggesting that pausing music lowers external noise. Later explanations of an independent loop and local cancellation inside the earcup are correct.
Loop, seal and operating limits · pass
The microphone → processor → speaker loop and passive/active roles are clear. Residual noise and difficulty fully removing speech are retained, with explicit limits on whole-cabin silence and measurement claims.
Caption timing and embedded subtitles · pass
All 15 SRT cues have valid times tied to real TTS segment samples. Text and timecodes extracted from the MP4 mov_text stream match the original SRT.
Measured narration and mix · partial
The 86.15-second Kokoro voice matches text/audio receipt hashes without tempo or pitch changes. The music/effects bed is about 13dB below voice RMS and the final soundtrack is padded to 88.15 seconds. No listening review was performed.

Task source

Uses the frozen November 4 topic and narration guide with the shared Kokoro tool. This brief is a summary; the hash identifies the original session prompt.

Bose: active cancellation, music independence, speech and seal limits

persistent-author-session · native default · model_finished · Requests / complete responses: 50 / 49 Run record

One persistent Hy4 session authored the work; observed usage is stored separately, and missing usage is not counted as zero.

Source brief and prompt
题面概要:依据当天冻结的Bose机制资料,解释不放音乐时主动降噪如何减少噪声,制作60–90秒、1920×1080中文口播、字幕与机制动画。
提示词概要:由Hy4依据冻结事实与口播指南写讲稿、读音稿、分镜,区分物理隔音/主动回路、展示残余和适用边界,交付中文旁白、同步字幕、成片与可复建源码。

SHA-256: a9ee792c3cf242e1c4116f43e8129dc1139a6a3092823571bee4f5250acfa533

SVG and animation

An analog clock at 9:41: render of unchanged author SVG

An analog clock at 9:41

Correct 9:41 hand angles and relative lengths, all 12 numbers, and a clear dark and gold dial with its time label.

Source projectSVG

Test details

6m 56s · Requests / complete responses: 22 / 22 · 428,673 / 33,109 input / output tokens

  • Hour and minute angles of 290.5° and 246°, with lengths of 113 and 150, match 9:41.
  • All twelve numbers are correctly placed; the 09:41 label is complete, with no second hand.
  • The static render has clear ticks, numerals and gold details, with no observed clipping.

Itemized verification

Self-contained static SVG · pass
final.svg passes XML and resource checks, uses viewBox 0 0 800 600, and contains no scripts, external resources, embedded images, SMIL or CSS animation.
9:41 hand geometry · pass
Actual path tips transformed by CTM measure 290.5° for the hour hand and 246° for the minute hand, clockwise from 12. Lengths of 113 and 150 make the hour hand shorter.
Numbers and second hand · pass
Numbers 1–12 are complete and correctly positioned; the 09:41 label sits below the dial. Only hour and minute hand groups are present, with no second hand.
Actual render and layout · pass
Inspected the 1200×900 PNG from the unchanged SVG. Numbers and hands are clear, with coherent gold trim and a dark background, and an unclipped label.

Task source

Uses visual-svg-v2 and creative-suite-v4. This is a summary; the prompt hash identifies the actual archived creation session.

persistent-author-session · native default · model_finished · Requests / complete responses: 22 / 22 Run record

Created and inspected in one persistent Hy4 session; the coordinator did not alter the original SVG.

Source brief and prompt
题面概要:静态SVG钟表显示9:41;1至12完整,时针短于分针,角度随分钟正确变化,不画秒针;无脚本、外链或嵌入图像。
提示词概要:静态SVG钟表显示9:41;1至12完整,时针短于分针,角度随分钟正确变化,不画秒针;无脚本、外链或嵌入图像。

SHA-256: f7df9c7d6ff8a4e59a5b977d970f8e95cd3662a055f5f7fed9eb23b37c420168

Data chart and proportional bars: render of unchanged author SVG

Data chart and proportional bars

All four bars share a zero baseline, with correct proportions, ticks, quarter labels and summaries in a clear light card layout.

Source projectSVG

Test details

4m 14s · Requests / complete responses: 13 / 13 · 187,907 / 20,033 input / output tokens

  • Actual bar heights are 120, 240, 180 and 300, in the ratio 2:4:3:5.
  • Q1–Q4 and bar values are present; Total: 84 and Best: Q4 are correct.
  • Colors and hierarchy are clear; inspection found no text collisions or clipping.

Itemized verification

Self-contained static SVG · pass
XML and resource checks pass; viewBox is 0 0 800 600, with no scripts, external resources, embedded images, SMIL or CSS animation.
Proportions and zero baseline · pass
Chromium measures actual path heights of 120, 240, 180 and 300, with the same baseline at y=470. They match values 12, 24, 18 and 30 in the ratio 2:4:3:5.
Labels and summaries · pass
Q1–Q4, bar values 12/24/18/30, axis ticks 0/10/20/30, Total: 84 and Best: Q4 are all present.
Actual render and layout · pass
Visually inspected the 1200×900 PNG from the unchanged SVG. The light card layout is clear, with no observed text collision, clipping or overflow.

Task source

Uses visual-svg-v2 and creative-suite-v4. This is a summary; the prompt hash identifies the actual archived creation session.

persistent-author-session · native default · model_finished · Requests / complete responses: 13 / 13 Run record

Created and inspected in one persistent Hy4 session; the coordinator did not alter the original SVG.

Source brief and prompt
题面概要:静态SVG绘制Q1–Q4的12/24/18/30四柱,零基线、线性刻度,标注Total:84和Best:Q4。
提示词概要:静态SVG绘制Q1–Q4的12/24/18/30四柱,零基线、线性刻度,标注Total:84和Best:Q4。

SHA-256: ee4989094580ee5393733ba96c781b5cc5ac7c8f4e8032d7e983bed0b1f81d20

Chinese poster and layout constraints: render of unchanged author SVG

Chinese poster and layout constraints

All six Chinese strings appear exactly once, with clear hierarchy for the headline, event details and registration button, plus a coherent bicycle icon and spacing.

Source projectSVG

Test details

9m 1s · Requests / complete responses: 27 / 27 · 1,036,148 / 47,544 input / output tokens

  • All six required strings match exactly, with clear Chinese glyphs and no missing or extra text.
  • The headline is largest, time/place follow, and the note is smallest; final text boxes remain inside the canvas without overlap.
  • A light background and orange/green palette support a recognizable bicycle and ample margins.

Itemized verification

Self-contained static SVG · pass
XML and resource checks pass; viewBox is 0 0 800 600, with no scripts, external resources, embedded images, SMIL or CSS animation.
Six exact text strings · pass
Each of the six required strings appears exactly once, without additions or rewrites; the SVG has six visible text elements.
Type hierarchy and bounds · pass
Font sizes are 64 for the headline, 32 for time/place, 26 for the subtitle, 22 for the button and 16 for the note. The note is smallest; all measured text boxes fit the canvas and do not overlap.
Rendered layout and bicycle icon · pass
Inspected the 1200×900 PNG and final Chromium render. Chinese text is clear, the orange bicycle has two wheels, frame, saddle and handlebar, and margins and spacing are ample.

Task source

Uses visual-svg-v2 and creative-suite-v4. Event details belong to a fixed task, not a real event announcement.

persistent-author-session · native default · model_finished · Requests / complete responses: 27 / 27 Run record

One persistent Hy4 session authored the work; run metadata and missing usage are recorded separately.

Source brief and prompt
题面概要:六段指定中文各出现一次,分清标题/活动信息/按钮/说明层级,配自行车图标,文字不得溢出或重叠;输出自包含静态SVG。
提示词概要:保留六段指定中文并做清楚的字体层级与版面布局,绘制自行车图标,交无外部资源的静态SVG。

SHA-256: 10e6d1254b19f45927d4b81a5c7fab80dee5869224b6d55f8996d39c95e55bff

Cycling pelican: four-second preview of unchanged author SMIL SVG

Animated cycling pelican

A recognizable pelican and coherent bicycle, with rotating wheels and crank, connected pedaling legs, and a visually smooth four-second loop.

Source projectSVG

Test details

21m 27s · Requests / complete responses: 54 / 54 · 3,935,938 / 93,850 input / output tokens

  • The long bill and throat pouch are recognizable; wheels, frame, saddle and handlebar are complete, and the subject stays in frame.
  • Spokes and pedals actually move; leg joints stay connected across 41 samples, with maximum pedal/crank endpoint error of about 0.183 native SVG units.
  • All 40 frames differ across the four-second loop. Live playback crosses a loop; the endpoints have a very small pixel difference with no obvious jump observed.

Itemized verification

Self-contained SMIL animation · pass
XML and resource checks pass. All 48 SMIL animations repeat every four seconds; there are no scripts, CSS animations, external resources/fonts or embedded images.
Pelican and bicycle structure · pass
Representative frames and live-playback screenshots show a recognizable long bill and orange throat pouch, two wheels, a connected frame, saddle and handlebar, with the subject inside the canvas.
Wheels, crank and pedaling · pass
Spokes and crank visibly rotate, with the legs following the pedals. Across 41 samples at 0.1-second intervals, hip/body and knee gaps are zero; ankle/pedal error stays below 0.011 and pedal/crank error below 0.183 native SVG units.
Four-second loop and live playback · pass
A four-second capture has 40 distinct frames and passes full MP4 decoding. A separate 5.37-second Chromium playback observes one video loop and continuous SMIL motion. Frames at 0/4 seconds are not pixel-identical (MAE 0.00535/255), with no obvious visual jump observed. Separate real-time screenshots confirm WebP playback, with 40 frames totaling 4,000 ms.

Task source

Uses visual-svg-v2 and creative-suite-v4. This is a summary; the prompt hash identifies the actual archived persistent creation session.

persistent-author-session · native default · model_finished · Requests / complete responses: 54 / 54 Run record

SVG authored and motion inspected in one persistent native Hy4 session; the coordinator only captured/encoded the unchanged source.

Source brief and prompt
题面概要:以自包含SMIL SVG绘制鹈鹕骑自行车,长喙与喉囊可辨,双轮、车架、车座与车把齐全;轮辐、曲柄与相连双腿实际运动,4秒循环,主体保持在画布内,无脚本、CSS动画、外部素材或嵌入图像。
提示词概要:以自包含SMIL SVG绘制鹈鹕骑自行车,长喙与喉囊可辨,双轮、车架、车座与车把齐全;轮辐、曲柄与相连双腿实际运动,4秒循环,主体保持在画布内,无脚本、CSS动画、外部素材或嵌入图像。

SHA-256: 03b6e4a78c9dbaed8ba97ba86422d3b24bc63fd223ebe596833b79c83864fed1

Download all creative results and source index

Run settings, usage and full notes

Tencent Hy4 · 2026-11-04

creative-suite-v4: three Blender tasks, four Chinese knowledge films using the frozen Shanghai November 4 topics, three static SVGs and one cycling animation. The shared production instructions require 3D main views of at least 1920×1080 and other views of at least 1600×900; original attachment 4K targets are assessed separately. Actual tencent/hy4 authored and reviewed work in persistent native Claude Code sessions. A coordinating AI independently inspected meshes, final media, text and animation, separating delivery from correctness.

Creative responses report a known usage.cost subtotal of $3.460043. This is not invoice reconciliation and does not use CLI dollar estimates. Vision cost is listed separately. Requests use the provider's 65,536-token output ceiling, without total request, token, wall-time or retry-count termination budgets.

Tasks ran concurrently and continued their original projects. Once readable previews or main deliverables existed, coordinator prompts focused the author on missing final production and required notes; artwork source was not hand-edited. For the interior, the coordinator ran the unchanged author encoder in an offline sandbox on 180 existing frames and audio. This is production execution, not new model authorship. Kokoro 0.9.4 / zf_xiaobei / z is the shared voice tool, not native Hy4 speech.

Request counts cover creative generation calls, excluding auxiliary count_tokens RPCs, readiness checks and the separate vision set. Input includes cache reads and output includes reasoning. 7 requests lack complete usage; tokens and cost are known subtotals, with unknown usage not treated as zero. Case wall times include tools, rendering, waiting and coordination gaps. Concurrent case times are not additive elapsed time or pure model latency.

Visual commentary by AI; dimensions and media checks use scripts.

Requests / complete responses: 771 / 764 · Errors / interruptions: 5 / 2

59,202,323 / 1,183,151 input / output tokens, including cache and thinking

  • Creative results come from persistent tool sessions rather than single first answers; only the separate vision set retains independent first complete responses.
  • A coordinating AI supplied visual observations and checks. Scripts do not prove every geometric detail, scientific statement or production fitness. Audio received technical and timeline checks, without a complete listening review.
  • Source and observed errors are preserved. Some rebuild scripts retain local tool/font paths; ZIP CRC and source inspection do not establish a full rebuild on another machine.
  • No actual desktop operation, natural-photo, long-scan or video-input test, and no matched Hy4 Preview comparison. Official benchmarks are separately attributed and were not reproduced.

Official benchmarks and conditions

Tencent-reported figures, not independently reproduced by TOKRACE. Comparisons use the highest available reasoning setting; tools, harnesses and budgets vary by benchmark.

ZeroBench-main is reported with tools at pass@5. BuddyVQA Atlas is an internal benchmark, with separate tool and no-tool results. OSWorld 2.1 uses Full, partial score; 78.9 is not a full-task completion percentage.

Benchmark-name asterisks mark first evaluations at release; model-cell asterisks in the appendix mark Tencent testing. Preview scores may change with harness, judge and anti-hacking updates. Read the original appendix for the complete conditions.

Official Hy4 benchmark overview: coding, agents, research and computer use
Official Hy4 benchmark overview: coding, agents, research and computer use · Tencent Hy / 腾讯混元 · Official source · Full original with axes, legends and footnotes
Official hard-benchmark gaps: research mathematics, long-horizon coding, foundation-model improvement and vision without tools
Official hard-benchmark gaps: research mathematics, long-horizon coding, foundation-model improvement and vision without tools · Tencent Hy / 腾讯混元 · Official source · Full original with axes, legends and footnotes
Official full benchmark appendix and methodology footnotes
Official full benchmark appendix and methodology footnotes · Tencent Hy / 腾讯混元 · Official source · Full original with axes, legends and footnotes

Evidence and downloads

Direct OpenRouter image input with no tools, external OCR, or answer key. Each task has an independent one-message request; the first complete response is retained. Provider-default temperature and a 65,536-token output limit. A coordinating AI scores 25 checks against ground truth frozen before responses.

Five tasks use six programmatically drawn PNGs, with no tools and the first complete response retained. The model passes all 25 fixed checks across a Chinese sheet, chart, floorplan, image pair, and explicitly instructed occlusion boundary. This small set does not cover natural photos, long real scans, video understanding, uninstructed hallucination stress tests, or a matched Hy4 Preview image-input comparison.

The five image-input requests report 10,643 input tokens and 26,245 completion tokens, including 24,673 reasoning tokens. OpenRouter reports $0.036885 in usage.cost; this has not been reconciled against an invoice.

Creative results and sourcesImage-input prompts, answers and usageOfficial image sources and hashes

Hy4 launch analysis · More model reports