Disappointment

#11
by Daemontatox - opened

K2-Think is a 32 billion parameter open-weights general reasoning model with strong performance in competitive mathematical problem solving.

Its really lacking on the Reasoning side , reminds me of the very first trained models for reasoning where it kept getting stuck on "but wait" or "let me think".

Evaluation & Performance

from the table , its clear you are claiming your model beats Qwen3-32B on most benchmarks and Qwen3-235B on some benchmarks , while in real testing and evaluation , this barely beats Qwen3-14B.

Inference Speed

I dont know why you are boasting this since its not your architecture that has the model working that fast or how you mapped the model , its just you are using Cerebras WSE-3 cards , so i am not sure why you taking credit or using this as a "Feature".

Conclusion :
you basically Badly Finetuned a Qwen2.5 model on Qwen3-235B traces and you managed to mess it up.

Sign up or log in to comment