Minimax M3 is a really good model and its better than deepseek V4 flash, qwen (under 400b) and have really surprisingly good coding ability. But sometimes output is inconsistent.
Posted by 9r4n4y@reddit | LocalLLaMA | View on Reddit | 13 comments
[I am most likely assuming it's under 500B due to pricing]
For those who don't know, yeah, Minimax M3 is out. If you don't have good hardware, you can try it online on Minimax Agent Chat.
So whenever a new model comes out, I test 2 main things:
🔻1. Web search - it outperforms DeepSeek V4 Flash in web searches as it pulls off better results and finds niche stuff.
It also beats Gemini 3.5 Flash for web search. Like, idk why Gemini can't find some niche info properly from the web (I think it only reads snippets, that's why maybe).
It also outperforms Qwen 3.5 122B and Qwen 27B in web searches.
🔻2. (Only frontend) My personal standalone HTML test (this includes some games, webpages, and some SVG stuff)
Ngl, I was totally shocked that M3 was able to do some of my harder HTML tasks. It was sometimes beating Sonnet 4.6, but I think it's most likely not better than Sonnet 4.6. I would say it can do 90% of the stuff.
It outperforms DeepSeek V4 Pro in coding stuff (frontend only).
It also outperforms all Qwen models from all the latest series under 400B.
I would say it would be totally banger if it's under 300B parameters.
🔻 Now the cons - the only problem with Minimax is that sometimes it gives very good output and sometimes totally shit output, but only for coding tasks.
Share your opinion also :)
dinerburgeryum@reddit
I've found MiniMax to be wildly inconsistent, personally. I tried driving 2.7 for a few days and it, surprisingly, hallucinated way more than Qwen 27B or Gemma 31B. Not super acted about the new version but willing to be surprised.
9r4n4y@reddit (OP)
Yeah inconsistencies r still here but not as high as 2.7Â
NoCucumber6245@reddit
M3 is amazing. Im addicted to building, im building 10x the speed than before. Its so cheap and so good.
ttkciar@reddit
Until such time M3 weights have been published, this is off-topic for LocalLLaMA.
Queasy-Contract9753@reddit
I'll try it out later when I get a minute, but to be fair minimax agent system on their site has always been S tier. Specially for search. Feels a bit unfair to judge the model itself on a criteria that depends on the framework.
silenceimpaired@reddit
I’ll try it out when it’s on Huggingface as a GGUF by Unsloth
Plappedudel@reddit
I really hope that pricing will stay at current levels. M3 seems like a big step up over M2.7 and while it's unlikely to be truly frontier, their coding plan has really generous limits. If they can keep it that way, it's easily worth it.
sn2006gy@reddit
The model hasn't even been out a day, too early to have opinions.
9r4n4y@reddit (OP)
Take ur time then ¯_(ツ)_/¯Â
nomorebuttsplz@reddit
wake me up when parameter count is known.
complexminded@reddit
The comparison to qwen 27b its inevitable someone will ask "but does it beat qwen 27b" because that's there only frame of reference for a "good" model.
9r4n4y@reddit (OP)
You got it ;)
LagOps91@reddit
Pricing can always be subsidized, so i wouldn't count on it being sub-500b. would be nice if it was.