The model suffers from ADHD
My impression is that the model suffers from ADHD. It is unable to formulate thoughts in a coherent manner, let alone generate a narrative: instead of a smooth narrative, the result is a somewhat chaotic collection of notes. However, compared to Ministral-3-14B-Base, this model is functional, as Ministral-3-14B-Base is completely dysfunctional.
It should be noted that this review is based on the performance of both models with German-language texts. It should be remembered that Mistral was previously truly multilingual. Unfortunately, this is no longer the case.
It suffers from corporatitis and compulsive gaslighting, like ALL of them.
It's making me suffer because of how many asterisks it uses. Sometimes two next to each other, sometimes even three. It doesn't even know what and why it is doing that I bet.
It's making me suffer because of how many asterisks it uses. Sometimes two next to each other, sometimes even three. It doesn't even know what and why it is doing that I bet.
Those asterisks are just markdown style (bullets, bold, italic etc). But yeah, if you use it in plain text, it looks like Morse code.
My impression is that the model suffers from ADHD. It is unable to formulate thoughts in a coherent manner, let alone generate a narrative: instead of a smooth narrative, the result is a somewhat chaotic collection of notes. However, compared to Ministral-3-14B-Base, this model is functional, as Ministral-3-14B-Base is completely dysfunctional.
It should be noted that this review is based on the performance of both models with German-language texts. It should be remembered that Mistral was previously truly multilingual. Unfortunately, this is no longer the case.
BTW, in case you have that option, did you choose a proper chat template? When I manually wrote tags and made small mistakes, output was literally insane but that was not the model's fault.
No I mean, it loves wrapping everything in multiple asterisks. I'm parsing the markdown properly but I feel like it's wasting tokens on unnecessary asterisks.
BTW, in case you have that option, did you choose a proper chat template? When I manually wrote tags and made small mistakes, output was literally insane but that was not the model's fault.
Yeah I should have. I'm currently using v7-tekken
I tried both chat and text completion and after switching to koboldcpp I tried different ggufs, from the original Q5_K_M to Q8_0. Even some exl3 quants . I tried the recent
:
mradermacher/Ministral-3-14B-Instruct-2512-absolute-heresy-GGUF · Hugging Face
At Q_8_0 as well. No luck. It keeps shoving so many asterisks into a response. Everything is bold, italics, and both at the same time.