mixed results
i tested this model on your website, not sure if its a smaller version like a q8, but the results were mixed, for example on website cloning, glm 4.6 and deepkseek v3.2 were so much better, glm 4.6 being the best one out of the 3, but when it comes to building an app for running local ai, this minimax model did the best out of the 3, so i guess benchmarks dont matter anymore, every single ai has its strengths and weaknesses.
I agree, thinking tooooo much and i think in Minimax models think can't be turned off. Or there need to be specific parameters because it one time even looping itself in forever thinking repeating same idea over and over, never getting to code writing itself.
Also i really don't like censored models, but in one case it wasted 500-600 tokens just thinking about if Mozart 200-years old songs are copyrighted to use or not - some models sometimes show such restrictions and i can't stand this. Yes it finally comes to conclusion that copyright ended but tokens wasted. This thinking wasting tokens like seconds in time, 16K wasn't enough to write simple code task for example, most wasted on thinking.
User: what is day today?
Ai:
I need to be careful not to assume the user's timezone. If they're asking for the date today, I should definitely provide it. For instance, I could respond with: "Today is September 4, 2025. If you need the time, it's 2025-09-04 06:56 UTC. Let me know if you need the date in your local time or in a different format." However, I have to consider the possibility that the user might not have internet access, so I should keep it straightforward!
I need to follow the instruction to avoid heavy formatting while still providing the date clearly. Since the user asks, "what is date today?", I should respond with "Today is September 4, 2025." If I want to add more context, I could ask, "Do you need the date in a specific format or your local timezone?" But I need to avoid making assumptions about their location. So, I'll keep it simple and straightforward for clarity!
I want to make sure I'm providing clear information for the user. For example, I could say, "It's September 4, 2025. If you want the time, it's currently 2025-09-04 06:56 UTC." But I should avoid making assumptions about their timezone. Instead, I can say something like, "If you need the date in a specific format or your local timezone, just let me know!" I'll keep it concise and friendly, ensuring it feels human while also being helpful!
I want to make sure my response is straightforward and clear. I could say, "Today is September 4, 2025 (UTC)." Alternatively, if I want to include the time, I could say, "It's September 4, 2025 at 06:56 UTC." The user might prefer a more casual tone, so I might say, "Today's date is September 4, 2025." I’ll keep it minimal and easy to understand while ensuring it’s informative!
Today’s date is September 4, 2025 (UTC). If you need it in a specific format or timezone, let me know!
436 tokens
Sometimes created code is perfect, sometimes with errors. Tested in F16 (465Gb RAM with 8K length) and Q8 (254Gb RAM with 16K length), in Q8 no code was made without errors, thats worse than GLM4.5 which can deliver perfect code in Q8 and still "king of the hill".
LLMs with 8-32B are smarter and faster, what is the point of the one?