Open-Source AI Bug Fixed: MiniMax-M3 Now Works Properly

If you've tried the MiniMax-M3 AI model recently and got gibberish back, you weren't alone. A bug in the code caused the model to produce random text—what developers call 'token soup.' The issue was that certain scaling factors (numbers that keep calculations in the right range) weren't being passed to the part of the model that handles expert layers. Without them, the model's outputs were thousands of times too large, drowning out the actual signal.

The fix, now merged into the open-source codebase, ensures those scaling factors are forwarded correctly at three key points. This matches how other similar models handle it. As a result, MiniMax-M3 should now generate coherent, useful responses. The bug was reported with a specific error code (G ~= 15123) and affected the M3.1 Flash version, but the fix applies broadly.

For anyone using MiniMax-M3—whether for research, building chatbots, or generating text—this means the model will finally work as expected. No more nonsense answers. The fix is a reminder that even small technical details can make or break an AI model's performance. It also highlights the collaborative nature of open-source development, where bugs are caught and fixed quickly by the community.

If you're not a developer, you might not interact with this directly, but it's good news for the AI tools you use. Many apps and services rely on open-source models like MiniMax-M3. When these models improve, the products you use get better too—more reliable, more accurate, and less likely to frustrate you with weird outputs. So next time you ask an AI a question and get a sensible answer, thank a bug fix like this one.