Hacker News Reader

The gap is huge and im tired of reading these articles constantly

Are you talking about hosted vs the ones you can easily run locally? Because there are open models that require hundreds of gb of vram which are apparently pretty close.

on the Will It Mythos benchmark, small models are punching way above their weight(s)

gemma4-26B (#7)

qwen-3.6-27B (#9)

https://news.ycombinator.com/item?id=48640196

I've tried running qwen 3.6 locally and it felt like LLMs a year ago where you can get them to do some stuff but the tasks have to be very small and you have to course correct them a lot to the point it's hard to say it's any faster than doing it all yourself.

Certainly the gap is closing but I feel it still makes more sense to pay pennies to run the full sized open models hosted on much better hardware.

I had qwen36moe revamp my PhD thesis with a rewrite using JAX. Gave it access to my old code, helpednitnwhen it got stuck or didn't quite understand a few times.

Overall I was very impressed with its open box reimplementation. I remain of the mind they are widely underrated.

What edition of Qwen 3.6? 35b-q6 (with MOE) has felt good enough for general purpose agentic coding.

You obviously need a 32GB card to run that (or a 64GB Mac), and realistically anything lesser than an RTX 5090 is going to be too slow for practical use.

Reply on news.ycombinator.com