Thanks alot!
This what I've been waiting for the reasoning version! Finally have all the versions of your guffs ha. thanks again!
Hey there @JanjanJean ,
Sorry for the wait! Lol, I had been struggling for a little with ensuring that the model would always utilize reasoning without errors in the output when responding. Currently I have found that with the system prompt I've embedded into the tokenizer chat template, finally, I can achieve this result and the model properly initiates tags while reasoning. It was an interesting "issue" to solve as the model seemed to be dead set on not utilizing reasoning and especially so with the standard system prompt from MistralAI. No bashing on MistralAI though - they released a solid model, just personally I found it required a little tweaking. I also have changed the generation settings for all three of the Ministral-Reasoning models I released. The settings have been altered to my personally used settings which resulted in adherence to the request for the model to reason, AND it resulted in fixed reasoning that did not end in the model looping forever as some have reported and it also corrected the model thinking forever without replying to the user.
Anyway, took much longer than I expected but so does everything in my life it seems ahahhaha. Glad you like it! (At least I hope you do). Keep an eye out for another release though that I'll be making very soon. It will be the Llama-3.3-8B-Reasoning (yes that's not a typo, 3.3-8B) that was leaked from Meta. I also have grafted in the Llama 3.2 Vision tower for the 11B which now makes the Llama-3.3-8B into an 11B vision model as well.
Anyway, I'm getting back to work - good to hear from you!