Genuinely Impressive Little Model!
Just wanted to say a massive thumbs up of appreciation to the team that worked on this. This little 1.5b CoT model has genuinely blown me, and my research partners away. I have found this model CONSISTENTLY out-performs R1 14b distill at FP16 for reasoning, speed, and math. I have tested tons of CoT models (Small thinker, QwQ, QvQ, OpenThinker 32b, , Sky T1, Marco o1, S1, S1.1, R 1.5b/7b/32b/70b distills, InternLM3, Small 3, DeepHermes3 8b, and more) and this model has been in several leagues of its own in terms of reasoning, especially at its size. It blows me away that it can maintain and effectively reason through sometimes over 20k token long CoT's, and end up with the right response.
I have verified it can solve: Calculus, linear algebra, differential equations, trigonometry, matrix multiplication, sigmoid functions, parabolas, and more, and it has basically a 100% accuracy (it has gotten over 32 Calculus questions correct 16/16 times each)
A friend of mine was trying to get an answer for a coding implementation, and when compared to o3 mini low, this tiny little local 1.5b CoT model, was on average more correct about the question. That's MENTAL!!!
This little 1.5b CoT model has become my go to for CoT use cases, even over R1 32B at 8bpw, simply because it is so damn good and fast. I can use it as a test model to try out some CoT questions before giving them to the 32B R1, but this model genuinely gets the questions right about 80% of the time, and I never even end up needing R1 32b most time.
I would kill to see what a 7b, 14b, or dare I say, even a 32b could do with this!
PS: I have made several exl2 quants of this model, and have found that it still maintains a consistent ability to do complex math all the way down to 3.25bpw. The 2.5bpw one can also sometimes solve math, but thats asking for a lot of it lol
Thank you for your love! We really appreciate it. We will continue to deliver great reasoning models and push the boundaries in this space!
Very excited to hear that you got the 1.5B working robustly on your end! Our goal is to release models that work for real and do solid work for the open-source community ;)