Let's Run MiniMax M3.1 Local AI - Better than DeepSeek & Opus? 🤔

Summary

The host tests MiniMax M3.1 Flash locally on a Mac Studio with an M3 Ultra and 512 GB of memory. Despite its “Flash” name, the model is described as roughly 250 GB in size, and the test run uses about 253 GB of RAM.

The host cites reported benchmark results that place the model near Opus on one benchmark and ahead of DeepSeek on another, while noting that scores depend on the evaluation harness.

The main focus is practical code generation. Across a range of single-prompt tests, M3.1 Flash often produces impressive, playable 3D scenes and applications, including a cat game, a helicopter simulator, a solar system, a volcano, a website, an infographic, and an anatomy visualization. Many generations run without runtime errors, though some require code fixes and others have visual or control issues.

Overall, the host considers the model unusually capable and versatile for local use, while noting that quality can vary with the settings and prompt.

Key Findings and Demonstrations

• 3D game generation: Cat-game prompts produce detailed environments, animations, sound effects, and playable mechanics. Higher-effort or adaptive runs sometimes add features such as stamina and collectibles. Some versions have issues with rotation, animation, or objects, and a few have runtime errors that the host fixes.
• Other creative coding: The helicopter demo includes detailed parts, assembly and disassembly, flight, landing, and wireframe views. Solar-system demos support planet selection and, in some versions, informational panels and speed controls. The host also tests a volcano, a website with widgets and graphs, an infographic, and a 3D anatomy scene.
• Image-to-game and revision tasks: Attempts to recreate a mobile game and a Tomb Raider-style character from a reference image have mixed results. Some parts work, but the generations miss requested mechanics or have character-model and control issues. Follow-up prompts improve some details.
• Reasoning tests: The model handles several simple logic and math questions, including trick questions that require challenging the premise. It is weaker on some factual or technical questions, such as a C++/store-subscription issue and a pop-culture lyric reference.
• Reliability: The host praises the amount of useful code produced in one response but emphasizes that “one-shot” generation is not guaranteed to be error-free. The most ambitious reasoning settings sometimes produce compile or runtime errors; corrections or follow-up prompts can help.

Settings and Practical Guidance

• The model uses an adaptive reasoning setup with effort levels such as low, medium, high, extra high, and max. The host compares these settings with reasoning disabled and with adaptive mode.
• Token counts vary considerably by task and setting. The cat-game generations range from roughly 24,000 to 53,000 tokens in the examples discussed, while simpler demos use fewer. Some runs also exceed a nominal 10,000-token limit.
• The host explains temperature as adding weighted randomness to token selection: lower temperature favors more predictable choices, while higher temperature can add variation but may introduce odd artifacts. The video describes MiniMax’s recommended combination as temperature 1 with top-p 0.95.
• For a dark volcano scene, the host suggests increasing the code’s hemisphere-light value. The video also notes that the model is supplied pre-quantized to NVFP4, with some higher-precision components, and advises care when quantizing it further.

Overall Assessment

M3.1 Flash delivers strong local code-generation results, especially for visual 3D prototypes, and offers a broad range of reasoning settings to experiment with. The demonstrations show substantial creative capability, but also inconsistent fidelity, occasional bugs, and some factual gaps. The host’s conclusion is positive: it is a compelling local model, provided users are prepared to inspect, run, and refine its output.

Main Speakers and Sources

• Main speaker: The video’s host from the xCreate channel, presenting and testing the model.
• External claims discussed: Reported MiniMax benchmark figures and comparisons with Opus and DeepSeek. The host presents these as benchmark claims, not as independently verified results from the demonstrations.