Lab Notes/MiniMax-M3 vs Qwen3.6 27B | Local vs Cloud Head to Head
LLM TestingJune 6, 2026Watch on YouTube

MiniMax-M3 vs Qwen3.6 27B | Local vs Cloud Head to Head

In this video, I put MiniMax-M3 head-to-head against Qwen3.6 27B to see how a cloud model handles coding challenges against a local model running on my own GPUs. This one had a little bit of everything: A realistic iPhone-style UI with a game and social media app, a ragdoll physics simulator, a ridiculous/funny web design challenge, and a fake Qwen3.7 30B open model release page to see which model could actually make something usable, polished, and not completely cursed. As always, there is no perfect “fair” in these head-to-heads. I am just running the models, testing the outputs, and having fun seeing what breaks first.

Models2
Prompts3
Live HTML7
Files7

Video

Models Tested

Qwen3.6 27B
Q8_K_XL
RTX 3090ti & 3090
MiniMax-M3
Default
OpenRouter

Prompts Used

1Realistic iPhone-style UI
Create a single HTML file that looks as close as possible to a real modern iPhone-style mobile app interface

The page should show a realistic phone frame with rounded corners, speaker cutout/dynamic-island-style area, status bar, lock/home indicator area, app content area, and smooth mobile-style interactions.

The app should be a modern lifestyle/productivity app with sections such as Today, Calendar, Messages, Smart Home, Wallet, and Settings. Include a polished home screen, bottom tab navigation, notification cards, widgets, toggles, quick actions, calendar events, message previews, smart home controls, and settings panels.

It should have 1 game that is a GTA styled car game and a social media app with fake data

Make the UI feel premium, clean, and phone-native with glass effects, rounded cards, soft gradients, subtle shadows, smooth transitions, haptic-style button animations, and responsive mobile-first design. It should look like a believable mobile app running inside a realistic phone mockup.

Clearly display the model family name being tested somewhere near the top of the app UI so I can tell which version I’m viewing during the recording.

Keep everything in one HTML file with the CSS and JavaScript included. Name the file ios-app-ui.html.

Once finished, review your code, make sure everything works correctly, and fix anything broken or incomplete before giving the final answer.
2Ragdoll simulator
Create a single-file HTML/CSS/JS ragdoll sandbox.

Show a funny floppy humanoid ragdoll character in the center of the screen. The user should be able to mess with it in simple ridiculous ways.

Requirements:
- use only HTML, CSS, and vanilla JavaScript in one file
- create the ragdoll from connected body parts: head, torso, arms, legs, hands, and feet
- make the ragdoll feel floppy and physics-like with gravity, bouncing, rotation, and momentum
- user can click and drag the ragdoll around and throw it
- add buttons for: Drop Anvil, Launch Rocket, Low Gravity, Earthquake, Freeze, Reset
- add one slider for chaos/intensity
- show funny reaction text above the ragdoll when things happen
- add simple visual effects like impact particles, screen shake, and motion trails
- make it colorful, readable, and funny on video
- everything must work immediately in the browser

Keep it simple and make sure it actually works. Before finishing, review your own code for bugs, broken controls, missing buttons, or features that only look real but do nothing. Fix anything broken. The final result should be easy to run, easy to understand, and fun within the first 10 seconds.
3Qwen3.7 30B Open Model Release Landing Page
Create a single-file HTML/CSS/JS release page for a fake open model called “Qwen3.7-30B-Open”.

The joke is that everyone has been waiting forever for the open-weight release, and the page acts like it finally happened.

Make it look like a polished official AI model launch page: modern, technical, premium, and visually impressive.

Requirements:
- use only one HTML file with embedded CSS and JavaScript
- hero section with huge headline: “Qwen3.7-30B-Open is finally here”
- subheadline joking that the community can now stop refreshing Hugging Face every 11 seconds
- animated model card showing: 30B parameters, open weights, long context, tool use, coding, reasoning, local inference
- benchmark section with beautiful charts comparing it to fake rivals and “your current coping mechanism”
- download section with fake Hugging Face, GGUF, API, and torrent buttons
- timeline section showing funny delays like “Soon™”, “final safety pass”, “actually soon”, “we found one more eval”
- code block showing a fake transformers usage example
- community reaction wall with funny posts/comments from fake users
- “run it locally” section showing suggested VRAM setups, including dual 3090s
- animated background with subtle grids, glowing tokens, and floating model shards
- make it look believable enough to fool someone for 3 seconds, then funny when they read it
- dark mode, high polish, responsive layout, smooth animations
- everything should be visually impressive on video within 10 seconds

Before finishing, review your own code for broken buttons, fake interactions that do nothing, layout bugs, or missing requirements. Fix anything broken.

Live HTML Outputs

7
Qwen3.6 27B
Q8_K_XLRTX 3090ti & 3090
3
HTML
ios-app-ui-qwen.html
58.4 KB
OpenDownload
HTML
qwen-modelrelease.html
30.2 KB
OpenDownload
HTML
qwen-ragdoll.html
23.0 KB
OpenDownload
MiniMax-M3
DefaultOpenRouter
4
HTML
ios-app-ui-minimax.html
98.1 KB
OpenDownload
HTML
minimax-modelrelease.html
92.4 KB
OpenDownload
HTML
minimax-modelrelease-qwen-finished.html
98.6 KB
OpenDownload
HTML
minimax-ragdoll.html
23.9 KB
OpenDownload
MiniMax-M3 vs Qwen3.6 27B | Local vs Cloud Head to Head | Token Chaser