3 ms·
I don't know why, but the benchmarks still fails to cover the difference between large models and small ones. The small ones are great for many things, includin
by Squarex 22d ago
I don't know why, but the benchmarks still fails to cover the difference between large models and small ones. The small ones are great for many things, including general coding, but the larger ones, like fable and astra, have some kind of intelligence that is not present in the small ones.
- sinuhe69 22d agoMore parameters = more facts stored. Knowledges are almost incompressible, where strong reasoning only requires a 3B core or so.
- yorwba 22d agoWeibo's VibeThinker manages with half of that: https://arxiv.org/abs/2511.06221 https://arxiv.org/abs/2511.06221 (They finetuned Qwen2.5-Math-1.5B for reasoning.)