Tech • AI • Robotics • Game
I Tested MiniMax M3.1 Flash: Coding, Gaming and Business!
5/10TL;DR
MiniMax M 3.1 Flash appears highly competitive on price and raw speed, but tests suggest it still trails leading US models in coding, complex generation and reliability.
KEY POINTS
Low-cost, high-speed positioning
MiniMax M 3.1 Flash is positioned as a budget-friendly flash model focused on efficiency rather than frontier-level intelligence. It was cited at $0.22 per million tokens and an output speed of about 116 tokens per second, roughly double a common range of 50 to 60. On one task-cost comparison, it reportedly completed work for about $283, versus roughly $2,500 for GLM 5.3 and $7,000 for Sonnet 5.5.
Behavior profile differs from rivals
The model was characterized as especially organized, but less transparent in its answers and highly assertive. By contrast, Opus and Sonnet were described as warmer and more transparent while remaining similarly confident. Comparisons with Qwen Max, Llama, DeepSeek, GLM 4 and Kimi k1 highlighted how sharply model behavior can differ even when systems appear to target similar use cases.
Intelligence gap remains
An overall intelligence score of 43 was cited for MiniMax M 3.1 Flash, below current leading US models clustered around 52 to 53. In a narrower reasoning comparison, MiniMax 01-Flash was said to edge Qwen 2.5 by 43 to 41, but Qwen 2.5 was still presented as the stronger value choice overall because it is cheaper and more capable across more tasks. More advanced systems such as Sonnet 6.1 were described as nearly 10 points ahead in intelligence.
Weaknesses in PDFs and code
The clearest benchmark weakness was PDF comprehension, where MiniMax scored about 21% against an average closer to 30% for peers. Coding was another soft spot, with a cited score near 39%, while top US models were around 50%. On Terminal Bench for computer-use tasks, the model was placed at 26%, far below a US average near 60%.
Website cloning was convincing
In an e-commerce test, the model produced a close replica of a major US retail site under the rebrand Amazing.fr. The generated storefront included menus, product selectors, login elements and an add-to-cart flow, and it changed the branding successfully. However, some secondary images failed to load, certain categories returned incomplete product lists, and at least one product image crashed.
Copying over understanding
The strongest criticism of the e-commerce result was that the model appeared to reproduce visible structure and assets rather than deeply re-create site logic. Product descriptions, images and layout were copied effectively, but missing secondary media and uneven category behavior suggested brittle implementation. That makes it useful for rapid mockups, but riskier for original, production-grade builds.
3D generation lagged badly
A like-for-like prompt to generate an explorable 3D Paris with day-night switching, shadows and landmarks such as the Eiffel Tower, the Louvre and Notre-Dame produced a much weaker result than a competing US model. The MiniMax build offered a guided tour mode, but it consumed heavy memory, felt less polished and suffered navigation slowdowns. The rival output was described as smooth, detailed and fully navigable in real time.
Game creation exposed limits
In a game-building test, the gap widened. A competing model delivered a polished quest-based Paris environment with fluid movement and a working run mode after about 1 hour 30 minutes of coding. MiniMax finished a game build in about 1 hour 07 minutes, but the result contained texture issues, a floating character and broken movement, indicating that faster completion did not translate into usable output.
Speed depends on the task
Although marketed as fast, task-level testing showed that performance varied sharply by workload. The 3D Paris scene from a rival model was ready in 18 minutes 47 seconds, while MiniMax was still processing after nearly 1 hour 45 minutes, which likely contributed to the lag seen during testing. The model can therefore be fast in token throughput without being fast in end-to-end delivery for complex coding and rendering jobs.
CONCLUSION
MiniMax M 3.1 Flash stands out as a cheap, fast model that can produce credible web interfaces and solid general outputs. But current evidence suggests it remains behind top US systems when tasks demand robust coding, advanced 3D generation and dependable execution.
Ask a question