Short answer: we gave Qwen4 Max, Qwen3.8 Max, Kimi K3.1 Pro and DeepSeek V4 Pro 0813 the same four prompts for animated scenes in a single HTML file, each at its highest reasoning effort. All 16 files ran. Qwen4 Max took 647 to 715 seconds on every test, the narrowest spread of the four, and wrote the shortest file in three of the four tests. DeepSeek V4 Pro 0813 was the fastest writer and Kimi K3.1 Pro wrote the longest files. All four run on EmpirioLabs through one API.
Watch the comparison
Each model writes its code at its measured speed, then its file runs. The music was generated by ACE-Step on EmpirioLabs.
Specs at a glance
| Qwen4 Max | Qwen3.8 Max | Kimi K3.1 Pro | DeepSeek V4 Pro 0813 | |
|---|---|---|---|---|
| Released | November 16, 2026 | August 3, 2026 | October 23, 2026 | August 13, 2026 |
| Context window | 983,616 tokens | 983,616 tokens | 1,048,576 tokens | 1,000,000 tokens |
| Input | Text, images and video | Text, images and video | Text and images | Text |
| Maximum output | 262,144 tokens | 131,072 tokens | 131,072 tokens | 393,216 tokens |
| Output token price | Second lowest of the four | Second highest of the four | Highest of the four | Lowest of the four |
| Cached input | Discounted cache read | Input rate | Discounted cache read | Input rate |
| Setting in this test | Reasoning effort xhigh (its highest) | Reasoning effort max | Reasoning effort max | Reasoning effort max |
Rates are on the pricing page and on each model's page.
Test 1: weather app
Prompt: "Create a single self-contained HTML file (all CSS and JavaScript inline, no external libraries, no CDN links, no imports) that animates a polished weather app screen on a full-window <canvas> that fills any window size. Requirements: it cycles every few seconds through sunny, rainy, snowy and stormy weather with smooth transitions; the sky behind it changes to match, with animated sun rays, rain, snowflakes and lightning; the temperature counts to its new value; and an hourly forecast strip and a seven-day chart animate in, laid out with clean typography and spacing like a modern phone app. Smooth 60fps animation that runs on its own from the first frame and loops forever."
| Model | Reasoning effort | Time | Output tokens | Lines in the file |
|---|---|---|---|---|
| Qwen4 Max | xhigh | 711 seconds | 38,086 | 360 |
| Qwen3.8 Max | max | 1,211 seconds | 52,281 | 760 |
| Kimi K3.1 Pro | max | 746 seconds | 34,588 | 1,351 |
| DeepSeek V4 Pro 0813 | max | 383 seconds | 26,244 | 1,064 |
Test 2: starship HUD
Prompt: "Create a single self-contained HTML file (all CSS and JavaScript inline, no external libraries, no CDN links, no imports) that animates a starship cockpit heads-up display on a full-window <canvas> that fills any window size. Requirements: stars streaming past the window as the ship banks, an artificial horizon that tilts with it, a radar sweep with blinking contacts, targeting brackets that track and lock onto a ship flying across the view, and gauges for shields, speed and fuel that rise and fall, all in one consistent cyan and amber interface style with thin crisp lines and small readable labels. Smooth 60fps animation that runs on its own from the first frame and loops forever."
| Model | Reasoning effort | Time | Output tokens | Lines in the file |
|---|---|---|---|---|
| Qwen4 Max | xhigh | 667 seconds | 33,720 | 728 |
| Qwen3.8 Max | max | 1,165 seconds | 50,453 | 699 |
| Kimi K3.1 Pro | max | 597 seconds | 24,224 | 990 |
| DeepSeek V4 Pro 0813 | max | 172 seconds | 10,931 | 423 |
Test 3: split-flap departures board
Prompt: "Create a single self-contained HTML file (all CSS and JavaScript inline, no external libraries, no CDN links, no imports) that animates an airport split-flap departures board on a full-window <canvas> that fills any window size. Requirements: rows of flights with time, destination, flight number, gate and status for made-up airlines; characters flip through the alphabet one flap at a time with a mechanical bounce until they settle; every few seconds the top flight departs, the rows move up and a new one flips in; statuses change, with BOARDING blinking and DELAYED in amber; and an analog clock above with a sweeping second hand. Smooth 60fps animation that runs on its own from the first frame and loops forever."
| Model | Reasoning effort | Time | Output tokens | Lines in the file |
|---|---|---|---|---|
| Qwen4 Max | xhigh | 647 seconds | 37,289 | 450 |
| Qwen3.8 Max | max | 1,315 seconds | 56,231 | 543 |
| Kimi K3.1 Pro | max | 748 seconds | 28,789 | 959 |
| DeepSeek V4 Pro 0813 | max | 680 seconds | 43,083 | 1,260 |
Test 4: self-playing piano
Prompt: "Create a single self-contained HTML file (all CSS and JavaScript inline, no external libraries, no CDN links, no imports) that animates a self-playing piano on a full-window <canvas> that fills any window size. Requirements: a glossy keyboard of at least three octaves with correct black and white keys along the bottom, colorful note bars falling toward it in time with a cheerful melody and chords the program composes, each key pressing down and lighting up in its note's color while its bar touches it, a small burst of sparks on every hit, and the keyboard's light reflected on a dark polished floor. It is silent, with no audio and no start button. Smooth 60fps animation that runs on its own from the first frame and loops forever."
| Model | Reasoning effort | Time | Output tokens | Lines in the file |
|---|---|---|---|---|
| Qwen4 Max | xhigh | 715 seconds | 39,479 | 417 |
| Qwen3.8 Max | max | 741 seconds | 35,759 | 496 |
| Kimi K3.1 Pro | max | 770 seconds | 30,231 | 789 |
| DeepSeek V4 Pro 0813 | max | 468 seconds | 28,992 | 557 |
How we ran it
Each model received the identical prompt for each task as one user message, with no system prompt and no temperature, ending with the same instruction to return only the complete HTML file in one code block. Qwen4 Max ran at reasoning_effort xhigh, its highest level; Qwen3.8 Max, Kimi K3.1 Pro and DeepSeek V4 Pro 0813 ran at max. Every request streamed, with a 65,536 token output limit, on November 16, 2026. Times are wall-clock seconds for the whole request, and token counts come from the API's usage report. Each file is the first one the model returned for that prompt, rendered exactly as returned.
Three Qwen3.8 Max requests (the weather app, the starship HUD and the split-flap board) ended before returning a file and were sent again over WebSocket mode, which runs past the 15 minute limit of an HTTP request; their times are from the request that returned the file. A fifth test, a neon noodle bar, is not shown because Qwen4 Max and Qwen3.8 Max did not return a file for it.
What to look for
Time. Qwen4 Max took 647 to 715 seconds on every test, the narrowest spread of the four. Kimi K3.1 Pro took 597 to 770 seconds, DeepSeek V4 Pro 0813 172 to 680 and Qwen3.8 Max 741 to 1,315.
Output speed. Qwen4 Max wrote 50.5 to 57.6 tokens per second, Qwen3.8 Max 42.8 to 48.3, Kimi K3.1 Pro 38.5 to 46.3 and DeepSeek V4 Pro 0813 62.0 to 68.5.
File length. Kimi K3.1 Pro wrote the longest file in three tests and DeepSeek V4 Pro 0813 in the split-flap board. Qwen4 Max wrote the shortest file in the weather app, the split-flap board and the piano (360 to 450 lines). File length is not quality: the longest file is not always the scene that moves best.
The scenes. Compare each weather card's layout and transitions, each cockpit's radar and targeting, each board's flaps and clock, and each keyboard's falling notes and light. The reel shows the center of each full-window scene, so the outer edges of the wider layouts fall outside their panels.
Run the same test on EmpirioLabs
curl https://api.empiriolabs.ai/v1/chat/completions \
-H "Authorization: Bearer $EMPIRIOLABS_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "qwen4-max",
"reasoning_effort": "xhigh",
"stream": true,
"messages": [{"role": "user", "content": "Create a single self-contained HTML file (all CSS and JavaScript inline, no external libraries, no CDN links, no imports) that animates a self-playing piano on a full-window canvas. Return ONLY the complete HTML file inside a single html code block."}]
}'
Swap model to qwen3-8-max, kimi-k3-1-pro or deepseek-v4-pro-0813 (with reasoning_effort max) to run the same request on the others. A request like this takes minutes, so it streams the file as it is written. An HTTP request runs for up to 15 minutes; for a longer one, such as Qwen3.8 Max on three of these tests, send the same request over WebSocket mode on /v1/responses. Or try them in the Playground.
Frequently asked questions
Were the results edited or retried?
No file was edited or regenerated. Three Qwen3.8 Max requests that ended before returning a file were sent again, and a fifth test was dropped because Qwen4 Max and Qwen3.8 Max returned no file for it. Every panel is the first file its model returned, rendered exactly as returned.
Which setting did each model use?
Qwen4 Max ran at reasoning_effort xhigh, its highest level. Qwen3.8 Max, Kimi K3.1 Pro and DeepSeek V4 Pro 0813 ran at max.
Which model has the lowest rates?
DeepSeek V4 Pro 0813 has the lowest input and output rates of the four and Qwen4 Max the second lowest; Kimi K3.1 Pro has the highest output rate. Qwen4 Max and Kimi K3.1 Pro also bill cached input at a discounted rate. Each request costs what its tokens cost at the model's rates, so a model that reasons less or writes a shorter file costs less. See the pricing page for the numbers.
What is new in Qwen4 Max?
It is the next-generation flagship of Alibaba's Qwen family, released on November 16, 2026, for coding, frontend design, office work and complex agents, with image and video understanding, a 1M token context, three reasoning effort levels and five built-in tools. See How to Use the Qwen4 Max API.




