Is it just me or minimax-m2.7 is a regression in real world usage compared to minimax-2.5???

Posted by True_Requirement_891@reddit | LocalLLaMA | View on Reddit | 20 comments

I have been using the official api minimax-m2.7 and minimax-m2.5 in claude code since the first day of release and minimax-m2.5 always seems to complete tasks and figure things out faster than 2.7. Minimax-m2.7 halucinates too much, and I haven't see any improvement when it comes to real world usage in literally any task, but I have noticed regression. In terms of reliability 2.5 > 2.7 I have no idea why this is the case when it performs better on all benchmarks...

Reply to Post

20 Comments

[-]

Remper1997@reddit

If you are using the official on on Mac now you can track you api usage with this simple app: [https://github.com/Remper1997/MiniMaxUsage](https://github.com/Remper1997/MiniMaxUsage)

[-]

relmny@reddit

q4km is becoming my daily model in the PC that can run it. Used to be qwen3.5-27b... Time to test qwen3.6

[-]

mr_zerolith@reddit

What quant are you using?

[-]

My_Unbiased_Opinion@reddit

According to the UGI benchmark, 2.7 has a lower NatInt than 2.5. I find NatInt to be a VERY accurate general use benchmark. Your findings align with what I've seen as well.

[-]

crazyCalamari@reddit

Same here. Tried 2.7 on 3 projects to see if it lived up to the hype and the results were very underwhelming. Incorrect code, terrible native knowledge of solutions/framework (e.g. Temporal, Svelte, etc.), mediocre UI and unscalable architecture. Basically I had to redo all 3 for things I could even do with 120b models.

[-]

Mkengine@reddit

They really like to hype themselves to Sonnet levels, but at least M2.5 is more on par with Qwen3.5-35B-A3B than Sonnet 4.6: https://swe-rebench.com/

[-]

crazyCalamari@reddit

Interesting because Qwen 35B is what I would have put it with. It might try to mimic Sonnet style but in terms of intelligence I'm getting more from Qwen 122b and even Devstral 123b.

[-]

hainesk@reddit

Based on this it looks like people are really sleeping on Step-Fun Flash. Nearly the same as Qwen 397b.

[-]

LegacyRemaster@reddit

testing https://preview.redd.it/jj51g26riqug1.png?width=1864&format=png&auto=webp&s=5fcd802e4ea26f9dcc6e725374a499e4a1aa792f

[-]

True_Requirement_891@reddit (OP)

This is not just about the OS release. It just sucks in general on their official api.

[-]

RedParaglider@reddit

Then why are you bringing that up here instead of on the minimax forum?

[-]

LegacyRemaster@reddit

ah ok. today---> release on HF so yeah... You aren't talking about it. Locally is ok

[-]

LegacyRemaster@reddit

fix done. start C:\\llm\\llamaVulkan\\build\\bin\\Release\\llama-server.exe --model "E:\\Model\\unsloth\\MiniMax-M2.7-GGUF\\MiniMax-M2.7-UD-Q4\_K\_S-00001-of-00004.gguf" --ctx-size 81920 --threads 16 --host [127.0.0.1](http://127.0.0.1) \--no-mmap --jinja --fit on --flash-attn on --direct\_io -sm layer --n-cpu-moe 0 --threads 16 --cache-type-k q8\_0 --cache-type-v q8\_0 + cline + vscode.

[-]

MinimumCourage6807@reddit

Running local setup, few fast findings of unsloth ud-q3-xl (or whatever it is). 1. apparently there are some fixes to chat templates as either my agents or llamacpp are not constantly complaining about template problems (2.5 did, it worked but sometimes I could not see agent responces etc.). 2. 2.7 is a bit faster than about same sized 2.5 gguf. now the 2.7 runs in real world agentic loads around 80 t/s and pp around 2000 t/s. 2.5 was a bit slower in general. Might also be just updated llamacpp. 3. I havent got any tool call fails so far, 2.5 mess them a bit quite often, so I would say with few hours tests that 2.7 seems to be stronger in this. 4. The knowledge seems to be good and definitely the best i can run locally only in vram with 128 gigs by a big margin. Hard to tell yet is it better or worse than 2.5, as both does a good job. Hardware rtx pro 6000 + rtx 5090.

[-]

lowrizzle@reddit

I've had it go into full on loops a number of times in the last few weeks using it on openrouter. I get better results running Qwen3.5 35b locally, it's never stroked out on me like that. The last time, when I gave up, I let it loop 'let me build' over and over for about 15 minutes.

[-]

R_Duncan@reddit

check your quant, compare same quant of same author as often some quant aren't as good as others even being with the same size or name (i.e.: q4K\_M). For the 2.5 mxfp4\_moe is the one which performed better, here.

[-]

True_Requirement_891@reddit (OP)

read the post man...

[-]

Technical-Earth-3254@reddit

I've also had it hallucinate tools and classes more often than 2.5 (via API, can't run it locally). Idk if this was due to heavier demand at the time of usage (and them then using a lower quant mayb?).

[-]

Goldandsilverape99@reddit

I saw this regression from the minimax 2.1 to 2.5 also. and 2.7 is also bad. Both models can not solve the Chamber of Resonance puzzle in Indiana Jones and the Great Circle as a test question. The minimax 2.1 (and other models like stepfun-ai\_step-3.5-flash/mimo-v2-flash, and other models like qwen3.5-27B has no problem with this.

[-]

AppealSame4367@reddit

Do you use their coding plan? Because the one on their coding plan seems to not do well. When I used it via OpenRouter API when it was new, it was the real intelligence vs cost king, but I would prefer not to pay 10 cents for every request.