MiniMax M3.1 Flash vs DeepSeek V4.1 Flash 1017 vs Qwen3.8 Flash vs MiMo V2.6 Flash: four coding tests

Disclosure: This article was written with AI assistance and reviewed by EmpirioLabs AI.

Short answer: we gave MiniMax M3.1 Flash, DeepSeek V4.1 Flash 1017, Qwen3.8 Flash and MiMo V2.6 Flash the same four prompts for animated scenes in a single HTML file, one shot each. In our runs, MiniMax M3.1 Flash returned its file first in all four tests; MiMo V2.6 Flash cost the least on two of them, and MiniMax M3.1 Flash and Qwen3.8 Flash on one each. Fifteen of the sixteen files animate on their own; one stops at a syntax error. All four models are available on EmpirioLabs with the same API. We are not declaring a winner: watch the results and compare.

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. Watch it on YouTube.

Specs at a glance

Test 1: lighthouse in a storm

Prompt: "Create a single self-contained HTML file (all CSS and JavaScript inline, no external libraries, no CDN links, no imports) that animates a lighthouse on a rocky cliff on a stormy night on a full-window <canvas>. Requirements: the lamp at the top turns and its bright beam sweeps across the sea and through the falling rain, dark waves roll in and break into white foam against the rocks, slanted rain falls in layers, heavy clouds drift, and lightning flashes now and then light up the whole scene. Smooth 60fps animation that runs on its own and loops forever."

Test 2: bees building a honeycomb

Prompt: "Create a single self-contained HTML file (all CSS and JavaScript inline, no external libraries, no CDN links, no imports) that animates bees building a honeycomb on a full-window <canvas>. Requirements: a honeycomb of hexagonal wax cells grows outward across the screen, cells fill with glowing golden honey and then get capped with wax, a few dozen bees with flapping wings fly between flowers at the edges and the comb, and warm sunlight flickers over the scene. Smooth 60fps animation that runs on its own and loops forever."

Qwen3.8 Flash's file stops at a syntax error on line 329 of its script (2R where 2*R was meant), so the page shows only its background color. That is the file as returned, unedited.

Test 3: self-playing bowling lane

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 bowling lane seen from behind the bowler on a full-window <canvas>. Requirements: a glossy wooden lane in perspective, the ball rolls down it with a slight curve, hits ten pins that scatter and tumble with simple physics, the fallen pins are swept away and reset, and a scoreboard counts frames and strikes. Smooth 60fps animation that runs on its own and loops forever."

Test 4: self-playing claw machine

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 arcade claw machine on a full-window <canvas>. Requirements: a glass cabinet full of colorful plush toys in a pile, the claw moves over the pile, drops on its cable, closes around a toy, lifts it and carries it to the prize chute and drops it in, sometimes losing its grip so the toy falls back, blinking cabinet lights, and a counter of prizes won. Smooth 60fps animation that runs on its own and loops forever."

How we ran it

Each model received the identical prompt for each task, ending with the same instruction to return only the complete HTML file in one code block, and answered once. MiniMax M3.1 Flash, DeepSeek V4.1 Flash 1017 and Qwen3.8 Flash ran at reasoning_effort max; MiMo V2.6 Flash has no effort setting and ran with thinking on. Every request streamed, with a 65,536 token output limit, on October 19, 2026. Times are wall-clock seconds for the whole request, token counts come from the API's usage report, and prices are the amounts charged for those exact requests. Each returned file was recorded running in a headless browser, exactly as returned. We ran two more tasks in the same session, a rocket booster landing and a latte art pour; they are not part of this video.

What to look for

Time. MiniMax M3.1 Flash finished each test in 254 to 432 seconds, the shortest time in all four. Qwen3.8 Flash and MiMo V2.6 Flash took the longest on the lighthouse and the bowling lane.

Price. Each request cost between $0.009 and $0.076. DeepSeek V4.1 Flash 1017 has the highest output price of the four, so its requests cost the most in every test. MiMo V2.6 Flash cost the least on the bowling lane and the claw machine, MiniMax M3.1 Flash on the lighthouse, and Qwen3.8 Flash on the honeycomb, its shortest file.

The scenes. Compare each model's storm and beam, comb and honey, lane and pins, and claw and cabinet. File length is not quality: the longest file is not always the scene that moves best.

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": "minimax-m3-1-flash",
    "reasoning_effort": "max",
    "max_tokens": 65536,
    "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 arcade claw machine on a full-window <canvas>. Requirements: a glass cabinet full of colorful plush toys in a pile, the claw moves over the pile, drops on its cable, closes around a toy, lifts it and carries it to the prize chute and drops it in, sometimes losing its grip so the toy falls back, blinking cabinet lights, and a counter of prizes won. Smooth 60fps animation that runs on its own and loops forever. Return ONLY the complete HTML file inside a single ```html code block."}]
  }'

Swap model to deepseek-v4-1-flash-1017, qwen3-8-flash or mimo-v2-6-flash to run the same request on the others (for MiMo V2.6 Flash, replace reasoning_effort with "enable_thinking": true). A request this long can take several minutes; set "stream": true to receive the file as it is written. Or try them in the playground.

Frequently asked questions

Were the results edited or retried?

No. Every file in the video and the tables is the model's first answer to that prompt, rendered exactly as returned.

Why does one honeycomb show only a background?

Qwen3.8 Flash's file has a syntax error in its script, so the animation never starts. It is shown as returned, like every other result.

Which model costs the least per task?

In these four tests, MiMo V2.6 Flash on two, and MiniMax M3.1 Flash and Qwen3.8 Flash on one each. Each request costs what its tokens cost at the model's rates, so a model that writes a shorter file or reasons less costs less.

What is new in MiniMax M3.1 Flash?

It is MiniMax's new Flash model for coding and agentic work, released on October 19, 2026, with a 1M token context, image and video input, five reasoning effort levels and built-in web search. See How to Use the MiniMax M3.1 Flash API.

Try it

Playground | All models | Pricing